218alecturenotes - 6 07 12

MAT218: Lecture Notes on Partial Differential
Equations
Steve Shkoller
Department of Mathematics
University of California
Davis, CA 95616
June 7, 2012
Contents
1 Introduction 1
1.1 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Examples of PDEs . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2.1 Equation of continuity . . . . . . . . . . . . . . . . . . 2
1.2.2 The heat equation . . . . . . . . . . . . . . . . . . . . 3
1.2.3 The Laplace equation . . . . . . . . . . . . . . . . . . 4
1.2.4 The 1-d wave equation . . . . . . . . . . . . . . . . . . 4
1.2.5 Equilibrium of a membrane . . . . . . . . . . . . . . . 6
1.2.6 Equation for vibrating membrane . . . . . . . . . . . . 8
1.2.7 The Euler equations . . . . . . . . . . . . . . . . . . . 8
1.2.8 Minimal surface . . . . . . . . . . . . . . . . . . . . . 10
2 Introduction to Sobolev Spaces 11

2.1 Lp spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.1.2 Definitions and basic properties . . . . . . . . . . . . . 12
2.1.3 Basic inequalities . . . . . . . . . . . . . . . . . . . . . 13
2.1.4 The space (Lp (X), k · kLp (X) is complete . . . . . . . 15
2.1.5 Convergence criteria for Lp functions . . . . . . . . . . 16
2.1.6 The space L∞ (X) . . . . . . . . . . . . . . . . . . . . 17
2.1.7 Approximation of Lp (X) by simple functions . . . . . 19
2.1.8 Approximation of Lp (Ω) by continuous functions . . . 19
2.1.9 Approximation of Lp (Ω) by smooth functions . . . . . 20
2.1.10 Continuous linear functionals on Lp (X) . . . . . . . . 22
2.1.11 A theorem of F. Riesz . . . . . . . . . . . . . . . . . . 23
2
CONTENTS 3
2.1.12 Weak convergence . . . . . . . . . . . . . . . . . . . . 26

2.1.13 Integral operators . . . . . . . . . . . . . . . . . . . . 29
2.1.14 Appendix 1: The monotone and dominated conver-
gence theorems and Fatou’s lemma . . . . . . . . . . . 31
2.1.15 Appendix 2: The Fubini and Tonelli Theorems . . . . 32
2.1.16 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 The Sobolev spaces H k (Ω) for integers k ≥ 0 . . . . . . . . . 36
2.2.1 Weak derivatives . . . . . . . . . . . . . . . . . . . . . 36
2.2.2 Definition of Sobolev Spaces . . . . . . . . . . . . . . . 38
2.2.3 A simple version of the Sobolev embedding theorem . 40
2.2.4 Approximation of W k,p (Ω) by smooth functions . . . . 41
2.2.5 Hölder Spaces . . . . . . . . . . . . . . . . . . . . . . . 42
2.2.6 Morrey’s inequality . . . . . . . . . . . . . . . . . . . . 43
2.2.7 The Gagliardo-Nirenberg-Sobolev inequality . . . . . . 48
2.2.8 Local coordinates near ∂Ω . . . . . . . . . . . . . . . . 53
2.2.9 Sobolev extensions and traces. . . . . . . . . . . . . . 54
2.2.10 The subspace W01,p (Ω) . . . . . . . . . . . . . . . . . . 56
2.2.11 Weak solutions to Dirichlet’s problem . . . . . . . . . 58
2.2.12 Strong compactness . . . . . . . . . . . . . . . . . . . 60
2.2.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 64
2.3 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . 68
2.3.1 Fourier transform on L1 (Rn ) and the space S(Rn ) . . 68
2.3.2 The topology on S(Rn ) and tempered distributions . . 72
2.3.3 Fourier transform on S 0 (Rn ) . . . . . . . . . . . . . . . 73
2.3.4 The Fourier transform on L2 (Rn ) . . . . . . . . . . . . 74
2.3.5 Bounds for the Fourier transform on Lp (Rn ) . . . . . . 75
2.3.6 Convolution and the Fourier transform . . . . . . . . . 76
2.3.7 An explicit computation with the Fourier Transform . 77
2.3.8 Applications to the Poisson, Heat, and Wave equations 79
2.3.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 84
2.4 The Sobolev Spaces H s (Rn ), s ∈ R . . . . . . . . . . . . . . . 93
2.4.1 H s (Rn ) via the Fourier Transform . . . . . . . . . . . 93
4 CONTENTS
2.4.2 Fractional-order Sobolev spaces via difference quotient

norms . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
2.5 Fractional-order Sobolev spaces on domains with boundary . 103
2.5.1 The space H s (Rn+ ) . . . . . . . . . . . . . . . . . . . . 103
2.5.2 The Sobolev space H s (Ω) . . . . . . . . . . . . . . . . 104
2.6 The Sobolev Spaces H s (Tn ), s ∈ R . . . . . . . . . . . . . . . 106
2.6.1 The Fourier Series: Revisited . . . . . . . . . . . . . . 106
2.6.2 The Poisson Integral Formula and the Laplace operator109
2.6.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 112
2.7 Regularity of the Laplacian on Ω . . . . . . . . . . . . . . . . 113
3 Fourier Series and its Applications 119

3.1 Inner product spaces and Hilbert spaces . . . . . . . . . . . . 119
3.2 The Hilbert space L2 (T) . . . . . . . . . . . . . . . . . . . . . 122
3.2.1 Approximations of the identity . . . . . . . . . . . . . 122
3.2.2 Trigonometric polynomials . . . . . . . . . . . . . . . 124
3.2.3 Fourier representation of functions on [0, π] . . . . . . 126
3.3 Pointwise and uniform convergence of the Fourier series . . . 127
3.3.1 Pointwise convergence . . . . . . . . . . . . . . . . . . 128
3.3.2 Uniform convergence . . . . . . . . . . . . . . . . . . . 129
3.3.3 Jump discontiunity and Gibbs phenomenon . . . . . . 135
3.4 The Sobolev space H s (T) . . . . . . . . . . . . . . . . . . . . 136
3.4.1 Characterization of H 1 (T) . . . . . . . . . . . . . . . . 138
3.4.2 The Fourier representation of f 0 for f ∈ H 1 (T) . . . . 139
3.5 1-D heat equations with periodic boundary condition . . . . . 141
3.5.1 Formal approach . . . . . . . . . . . . . . . . . . . . . 141
3.5.2 Rigorous approach . . . . . . . . . . . . . . . . . . . . 142
3.5.3 The special case f = 0 . . . . . . . . . . . . . . . . . . 147
3.5.4 Decay estimates . . . . . . . . . . . . . . . . . . . . . 148
3.6 1-D heat equations with Dirichlet boundary condition . . . . 149
3.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
4 The Laplace and Poisson Equations 155

4.1 The fundamental solution . . . . . . . . . . . . . . . . . . . . 155
CONTENTS 5
4.1.1 Uniform and Hölder Continuous Functions . . . . . . 157

4.1.2 The Poisson equation −∆u = f in Rn . . . . . . . . . 159
4.2 Green’s function . . . . . . . . . . . . . . . . . . . . . . . . . 163
4.3 Properties of harmonic functions . . . . . . . . . . . . . . . . 165
4.3.1 Mean-value property . . . . . . . . . . . . . . . . . . . 165
4.3.2 Maximum principles . . . . . . . . . . . . . . . . . . . 166
4.3.3 The Harnack inequality . . . . . . . . . . . . . . . . . 167
4.3.4 Local estimates . . . . . . . . . . . . . . . . . . . . . . 168
4.3.5 Regularity of weakly harmonic functions . . . . . . . . 168
4.3.6 Liouville’s Theorem . . . . . . . . . . . . . . . . . . . 169
4.3.7 Analyticity . . . . . . . . . . . . . . . . . . . . . . . . 170
4.4 Computing Green’s functions . . . . . . . . . . . . . . . . . . 170
4.4.1 The case Ω = Rn−1 × R+ . . . . . . . . . . . . . . . . 171
4.4.2 The case Ω = B(0, 1) or Ω = B(0, R) . . . . . . . . . . 173
4.5 Perron’s method and solutions to the Poisson problem . . . . 174
4.6 Exercise . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
5 Second Order Linear Elliptic Equations 181

5.1 Strong solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 181
5.1.1 The Dirichlet problem . . . . . . . . . . . . . . . . . . 183
5.1.2 The Neumann problem . . . . . . . . . . . . . . . . . 183
5.1.3 The Robin problem . . . . . . . . . . . . . . . . . . . 183
5.1.4 A non-standard elliptic problem . . . . . . . . . . . . 184
5.1.5 A vector-valued Dirichlet problem: Linear Elasticity . 184
5.2 Weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 186
5.2.4 A non-standard elliptic problem . . . . . . . . . . . . 190
5.2.5 Linear elasticity . . . . . . . . . . . . . . . . . . . . . 191
5.3 The Lax-Milgram theorem . . . . . . . . . . . . . . . . . . . . 191
5.4 The BBL inf-sup condition . . . . . . . . . . . . . . . . . . . 193
5.5 Existence and uniqueness of weak solutions . . . . . . . . . . 197
5.5.1 A few Preliminary Lemmas . . . . . . . . . . . . . . . 197
6 CONTENTS

5.5.5 The non-standard elliptic problem . . . . . . . . . . . 203
5.5.6 The elasticity equations . . . . . . . . . . . . . . . . . 204
5.6 Weak solutions with distributional forcing functions . . . . . 206
5.7 General second order elliptic operators . . . . . . . . . . . . . 211
5.8 Horizontal convolution-by-layers and commutation estimates 215
5.8.1 Horizontal convolution-by-layers . . . . . . . . . . . . 215
5.8.2 Commutation estimates . . . . . . . . . . . . . . . . . 216
5.9 Elliptic regularity . . . . . . . . . . . . . . . . . . . . . . . . . 217
5.9.1 Dirichlet problem with homogeneous boundary condi-
tion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
5.9.2 Dirichlet problem with inhomogeneous boundary con-
dition . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
5.9.3 Neumann problem . . . . . . . . . . . . . . . . . . . . 223
5.9.4 The presence of b and c in the elliptic operator L . . . 224
5.9.5 Higher regularity . . . . . . . . . . . . . . . . . . . . . 225
5.10 Two-phase problems - Elliptic equations with mixed type of
boundary condition . . . . . . . . . . . . . . . . . . . . . . . . 226
5.10.1 The variational formulation and weak solutions . . . . 228
5.10.2 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . 229
5.11 Maximum principle . . . . . . . . . . . . . . . . . . . . . . . . 229
5.12 Eigenvalues and eigenfunctions . . . . . . . . . . . . . . . . . 231
5.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
6 Linear Parabolic Equations 241

6.1 Spaces involving time . . . . . . . . . . . . . . . . . . . . . . 241
6.1.1 Compactness theorems . . . . . . . . . . . . . . . . . . 245
6.2 Second-order parabolic equations . . . . . . . . . . . . . . . . 247
6.2.1 Weak solutions . . . . . . . . . . . . . . . . . . . . . . 248
6.2.2 Existence of weak solutions - Galerkin schemes . . . . 249
6.2.3 Uniqueness of the weak solution . . . . . . . . . . . . 255
6.2.4 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . 256
CONTENTS 7
6.2.5 Higher regularity and compatibility conditions . . . . 259

6.3 Parabolic equations with special forcing functions . . . . . . . 264
6.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
7 Linear Hyperbolic Equations 270

7.1 Weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . 270
7.1.1 Uniqueness of the weak solution . . . . . . . . . . . . 273
7.2 The method of parabolic regularization - the simple case . . . 275
7.2.1 The existence of the weak solution . . . . . . . . . . . 275
7.2.2 Regularity theory with smooth data . . . . . . . . . . 281
7.2.3 Regularity theory with data possessing limited regularity291
7.3 The parabolic regularization of general hyperbolic equations . 303
7.3.1 Estimates for the regularized quantities . . . . . . . . 306
7.3.2 Main theorem . . . . . . . . . . . . . . . . . . . . . . . 314
7.4 An alternative approach - the Galerkin method . . . . . . . . 315
7.4.1 Existence of weak solutions . . . . . . . . . . . . . . . 315
7.4.2 Regularity theory revisited . . . . . . . . . . . . . . . 319
7.5 Hyperbolic equations with Dirichlet boundary condition . . . 326
7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
8 More Topics on Elliptic Equations 332

8.1 Elliptic equations with Sobolev class coefficients . . . . . . . . 332
8.2 Elliptic decomposition using the divergence and curl . . . . . 336
8.2.1 An H 1 (Ω) elliptic estimate . . . . . . . . . . . . . . . 336
8.2.2 An H 2 (Ω) elliptic estimate . . . . . . . . . . . . . . . 338
8.3 The problem of divu = f in Ω with u = 0 on ∂Ω . . . . . . . 339
8.3.1 Inequalities associated with the normal and tangential
decomposition of vector fields on ∂Ω . . . . . . . . . . 341
8.4 Vector fields with given vorticity, divergence, and normal trace348
8.4.1 The case Ω = R3+ . . . . . . . . . . . . . . . . . . . . . 348
8.4.2 The case Ω is a bounded smooth domain . . . . . . . 350
8.5 Compensated compactness and the div-curl lemma . . . . . . 353
8.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
8 CONTENTS
9 Fluid dynamics 358

9.1 The equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
9.2 Incompressible Euler equations on a fixed domain . . . . . . . 359
9.2.1 Flow on a bounded and fixed domain . . . . . . . . . 359
9.2.2 The space of divergence-free vector fields on Ω . . . . 360
9.2.3 The Lagrangian flow of a velocity field . . . . . . . . . 362
9.2.4 Existence and uniqueness of solutions of the incom-
pressible Euler equations on fixed and bounded domain365
9.3 The Stokes equations . . . . . . . . . . . . . . . . . . . . . . . 369
9.3.1 The steady state case . . . . . . . . . . . . . . . . . . 369
9.4 Incompressible Euler with free-boundary . . . . . . . . . . . . 385
9.4.1 The Eulerian description . . . . . . . . . . . . . . . . . 385
9.4.2 The Lagrangian vorticity equation . . . . . . . . . . . 388
9.4.3 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . 388
9.4.4 Properties of the cofactor matrix a, and a polynomial-
type inequality . . . . . . . . . . . . . . . . . . . . . . 389
9.4.5 Horizontal convolution-by-layers and commutation es-
timates . . . . . . . . . . . . . . . . . . . . . . . . . . 390
9.4.6 An asymptotically consistent κ-approximation of the
Euler equations . . . . . . . . . . . . . . . . . . . . . . 392
9.4.7 Construction of smooth solutions to κ-approximate
Euler equations . . . . . . . . . . . . . . . . . . . . . . 394
9.4.8 Asymptotic estimates which are independent of the
smooth parameter κ . . . . . . . . . . . . . . . . . . . 396
9.4.9 Proof of the Main Theorem . . . . . . . . . . . . . . . 408
9.4.10 Time of existence and bounds independent of κ and
existence of solutions to (9.58) . . . . . . . . . . . . . 408
9.4.11 The limit as κ → 0 . . . . . . . . . . . . . . . . . . . . 409
A A Brief Review on Analysis 411

A.1 Metric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 411
A.2 Linear spaces (Vector spaces) . . . . . . . . . . . . . . . . . . 412
A.3 Normed spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 413
A.4 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . 413
CONTENTS 9
A.5 Completion of a metric space . . . . . . . . . . . . . . . . . . 413

A.6 Bounded linear maps between normed linear spaces . . . . . . 414
A.7 Convolutions and integral operators . . . . . . . . . . . . . . 415
A.8 The dual space and weak topology . . . . . . . . . . . . . . . 416
A.9 Bounded linear functional on Lp (Ω) . . . . . . . . . . . . . . 417
B Important Topics in Functional Analysis 419

B.1 The Hahn-Banach theorem . . . . . . . . . . . . . . . . . . . 419
B.2 The open mapping and closed graph theorem . . . . . . . . . 421
B.3 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . 421
B.3.1 Symmetric operators on Hilbert Spaces . . . . . . . . 430
Chapter 1
Introduction
1.1 Preliminaries
Theorem 1.1 (Divergence Theorem). Let Ω ⊆ Rn be a Lipschitz do-

main, that is, ∂Ω locally is the graph of a Lipschitz function, and w =
(w1 , · · · , wn ) ∈ C 1 (Ω) with outward pointing normal N . Then
Z Z
divwdx = w · NdS .
Ω ∂Ω
Now suppose that f is a scalar C 1 -function, and N = (N1 , · · · , Nn ) . By

setting w = f ei , where ei is the unit vector pointing to the positive xi -axis,
then the divergence theorem implies
Z Z
fxi dx = f Ni dS .
Ω ∂Ω
Suppose further that f is the product of two C 1 -functions h and g , then the
equality above implies
Z Z Z
ghxi dx = ghNi dS − gxi hdx .
Ω ∂Ω Ω
This is the multi-dimensional version of integration by parts.
Let Ω be a domain for which the divergence theorem holds and let u ∈
C 2 (Ω) and v ∈ C 1 (Ω)-functions. Then we have Green’s first identity:
∂u
Z Z Z Z
Dv · Dudx + v∆udx = div(vDu)dx = v dS . (1.1)
Ω Ω Ω ∂Ω ∂N
Suppose v ∈ C 2 (Ω) as well. Interchanging u and v in (1.1) and forming the

difference of the two equalities, we obtain Green’s second identity:
∂u ∂v i
Z Z h
(v∆u − u∆v)dx = v −u dS (1.2)
Ω ∂Ω ∂N ∂N
1
2 CHAPTER 1. INTRODUCTION
1.2 Examples of PDEs

1.2.1 Equation of continuity
Let u be the concentration of some physical quantity (u = u(x, t)) in a

domain Ω ⊆ Rn , and let F be the flux of the quantity, that is, F · NdS is
the rate of the quantity that passes through an area dS in the direction N
(outward pointing) normal to dS . Then
d
Z Z Z
udx = − F · NdS + qdx for all U ⊆ Ω ,
dt U ∂U U
where q is the strength of sources for the quantity. If u is smooth, by the

divergence theorem,
Z Z Z h i
ut dx = (q − divF )dx ⇒ ut + divF − q dx = 0
U U U
for all Lipschitz domain U . We then obtain the equation of continuity

ut + divF = q .
1. The conservation of mass: let %(x, t) and u(x, t) denote the density
and the velocity of a fluid at point x at time t . Then the flux F = ρu ,
and the equation of continuity reads
%t + div(ρu) = 0 ∀x ∈ Ω,t ∈ R. (1.3)
2. The traffic flow: let Ω = R (be the highway), and u denote the car den-
sity (given in vehicles per unit length). Then the flux F is a function
of u with the property that
(a) F (u) = 0 if u = 0 or u > L ,
(b) F 0 (u) > 0 if u ∈ (0, umax ) , and
(c) F 0 (u) < 0 if u ∈ (umax , L) .
If F is differential, and F 0 (u) = c(u) . Then the equation of continuity

reads
ut + c(u)ux = q ∀x ∈ R,t ∈ R.
1.2. EXAMPLES OF PDES 3
1.2.2 The heat equation
Let u(t, x) defined on (0, T ] × Ω be the temperature of a material body at

point x ∈ Ω at time t ∈ (0, T ] , and c(x) , %(x) , k(x) be the specific heat,
density, and the inner thermal conductivity of the material body at x . Then
by the conservation of heat, for any open set U ⊆ Ω ,
d
Z Z
c(x)%(x)u(t, x)dx = k(x)Du · NdS , (1.4)
dt U ∂U
where N denotes the outward-pointing unit normal of U . Assume that u

is smooth, and U is a Lipschitz domain. By the divergence theorem, (1.4)
implies
Z Z
c(x)%(x)ut (t, x)dx = div(k(x)Du)dx .
U U
Since U is arbitrary, the equation above implies
c(x)%(x)ut (t, x) − div(k(x)Du(t, x)) = 0 ∀ x ∈ Ω , t ∈ (0, T ].
If k is constant, then
X ∂2u n
c%
ut = ∆u ≡ .
k ∂x2i
i=1
If furthermore c and % are constants, then after rescaling of time we have
ut = ∆u . (1.5)
This is the standard heat equation (the prototype equation of parabolic

equations).
We need complementary conditions to specify a particular solution of
(1.5):
1. Initial condition: u(0, x) = ϕ(x) , ϕ(x) is a given function.
2. Boundary condition: if Ω is bounded, some boundary condition of u

at x ∈ ∂Ω for all time have to be introduced by physical reason to
specify a unique solution.
(a) Dirichlet condition: u(t, x) = ψ(t, x) for all x ∈ ∂Ω , t ≥ 0 , where

ψ is a given function.
∂u
(b) Neumann condition: = g for all x ∈ ∂Ω , t ≥ 0 , where g is a
∂N
given function.
∂u
(c) Robin condition: + hu = g for all x ∈ ∂Ω , t ≥ 0 , where h
∂N
and g are given functions.
1.2.3 The Laplace equation
The stationary solution (meaning that the solution is time independent) of

heat equation satisfies ∆u = 0 (the prototype of elliptic equations). Three
type of problems:

∆u = 0 in Ω ,
1. The Dirichlet problem:
u=ψ on ∂Ω .
∆u = 0 in Ω,
(
2. The Neumann problem: ∂u
=g on ∂Ω .
∂N
∆u = 0 in Ω,
(
3. The Robin problem: ∂u
+ hu = g on ∂Ω .
∂N
1.2.4 The 1-d wave equation
1. From Hooke’s law: imagine an array of little weights of mass m inter-

connected with massless springs of length h , and the springs have a
stiffness of k .
k k
m m m
u(x − h) u(x) u(x + h)
Here u(x) measures the distance from the equilibrium of the mass
situated at x . The forces exerted on the mass m at the location x are:
∂2u
FNewton = ma = m (x, t)
∂t2
FHooke = k[u(x + h, t) − u(x, t)] − k[u(x, t) − u(x − h, t)]
= k[u(x + h, t) − 2u(x, t) + u(x − h, t)] .
If the array of weights consists of N weights spaced evenly over the

length L = N h of total mass M = N m , and the total stiffness of the
array K = k/N , then
∂2u KL2 u(x + h, t) − 2u(x, t) + u(x − h, t)

(x, t) = .
∂t2 M h2
Taking the limit N → ∞ , h → 0 (and assuming smoothness) we
obtain
utt (x, t) = c2 uxx (x, t) . (1.6)
(2) Equation of vibrating string: let u(x, t) measure the distance of a

string from its equilibrium.
α β
T2 String
T1 u(x) u(x + h)
x x+h
Assuming only motion in the vertical direction, the horizontal compo-

nent of tensions T1 and T2 have to be the same:
T1 cos α = T2 cos β ≈ T .
The difference of the vertical component of T1 and T2 induces the

motion in the vertical direction:
∂2u
m (x, t) ≈ T2 sin β − T1 sin α = (T2 cos β) tan β − (T1 cos α) tan α
∂t2
≈ [T (x + h)ux (x + h, t) − T (x)ux (x, t)] .
If µ is the density of the string, then m = µh ; hence
∂2u T (x + h)ux (x + h, t) − T (x)ux (x, t)

µ 2
(x, t) ≈ .
∂t h
Taking the limit h → 0 , we obtain
µutt (x, t) = [T (x)ux (x, t)]x . (1.7)
If there is an external forcing f acting on the string, then (1.7) becomes
µutt (x, t) = [T (x)ux (x, t)]x + f . (1.8)
If T is constant, then (1.8) reduces to

f
utt (x, t) = c2 uxx (x, t) + . (1.9)
µ

u(x, 0) = ϕ(x) ,
Initial conditions: , where ϕ and ψ are given functions.
ut (x, 0) = ψ(x) ,
Boundary conditions:
1. Vibration string with fixed ends: u(0, t) = u(L, t) = 0 .

∂u ∂u
2. Vibration string with free ends: (0, t) = (L, t) = 0 .
∂x ∂x
1.2.5 Equilibrium of a membrane
Let T denote the tension of a membrane, and z = u(x1 , x2 ) be the membrane

in equilibrium.
dS
dA
Ω
The work done by the inner elastic force is equal to

dS
Z
E= T − 1 dA .
Ω dA
In other words, to deform a membrane from its unforced equilibrium state

to a surface S requires the input of the energy shown above. Since the
membrane S is parametrized by (x1 , x2 , u(x1 , x2 )) for (x1 , x2 ) ∈ Ω ,
Z hp i
E= T 1 + |Du|2 − 1 dx .
Ω
Now suppose that the deformation of the membrane is very small, or more
precisely, |Du| 1 , then by the Taylor expansion,
T
Z
E≈ |Du|2 dx .
Ω 2
If the deformation of the membrane is due to a small external force f , then
the potential energy stored in the membrane is
T
Z h i
E(u) = − |Du|2 + f u dx .
Ω 2
A membrane in equilibrium state (under the action of force f ) will minimize
the potential energy.
For ϕ ∈ C 1 (Ω) , let
E(u + tϕ) − E(u)
Z h i
δE(u; ϕ) = lim = − T Du · Dϕ + f ϕ dx .
t→0 t Ω
If u ∈ C 2 (Ω) , we may integrate by parts to obtain

∂u
Z h i Z
δE(u; ϕ) = div(T Du) + f ϕdx − Tϕ dS .
Ω ∂Ω ∂N
1. Membrane fastened on the boundary: then u satisfies δE(u; ϕ) = 0 for
all ϕ ∈ Cc1 (Ω) and u = g on ∂Ω , that is,
Z h i
div(T Du) + f ϕdx = 0 ∀ ϕ ∈ Cc1 (Ω) .
Ω
Therefore, div(T Du) + f = 0 in Ω . So the problem becomes

div(T Du) + f = 0 in Ω ,
(D)
u=g on ∂Ω .
2. Membrane with free boundary: then δE(u; ϕ) = 0 for all ϕ ∈ C 1 (Ω) ,

that is,
∂u
Z h i Z
div(T Du) + f ϕdx − Tϕ dS = 0 ∀ ϕ ∈ C 1 (Ω) .
Ω ∂Ω ∂N
In particular, consider ϕ ∈ Cc1 (Ω) , then we still obtain
div(T Du) + f = 0 in Ω;
∂u ∂u
Z
hence dS = 0 for all ϕ ∈ C 1 (Ω) . Therefore,
Tϕ = 0 on
∂Ω ∂N ∂N
∂Ω , and the problem becomes

 div(T Du) + f = 0 in Ω ,
(N) ∂u
 =0 on ∂Ω .
∂N
1.2.6 Equation for vibrating membrane
Let T be the tension, % be the density, and f be the density of the external
force which may depend on x and t .
d’Alembert’s principle:
Z h i
− T Du · Dϕ + (f − %utt )ϕ dx = 0
Ω
for all ϕ compatible with the existence constraints. Therefore,
1. Membrane fastened on the boundary:



 %utt − div(T Du) = f in Ω × (0, T ] ,

u=g on ∂Ω × (0, T ] ,


u(x, 0) = g(x), ut (x, 0) = h(x) for all x ∈ Ω .

2. Membrane with free boundary:



 %utt − div(T Du) = f in Ω × (0, T ] ,

∂u

=0 on ∂Ω × (0, T ] ,


 ∂N
u(x, 0) = g(x), ut (x, 0) = h(x) for all x ∈ Ω .

1.2.7 The Euler equations
Aside from the equation of continuity (1.3), at least an equation for the
velocity u is required to complete the system. Consider that conservation of
momentum. Let m = %u be the momentum. The conservation of momentum
states that
d
Z Z Z Z
mdx = − m(u · N)dS − pNdS + %f dx ,
dt U ∂U ∂U U
here we use the fact that the rate of change of momentum of a body is equal
to the resultant force acting onZthe body, and with p denoting the pressure
the buoyancy force is given by pNdS . Here we assume that the fluid is
∂U
invicid so that no friction force is presented in the fluid. Therefore, assuming
the smoothness of the variables, the divergence theorem implies that
n
∂(muj )
Z h X i
mt + + Dp − %f dx = 0 for all Lipschitz domain U ⊆ Ω .
U ∂xj
j=1
As a consequence, we obtain the momentum equation
(%u)t + div(%u ⊗ u) = −Dp + %f ∀x ∈ Ω,t ∈ R. (1.10)
Initial conditions: %(x, 0) = %0 (x) and u(x, 0) = u0 (x) for all x ∈ Ω .

Boundary condition: u · N = 0 on ∂Ω .
1. If the density is constant (such as water), then (1.3) and (1.10) reduce
to
ut + u · Du = −Dp + f in Ω × (0, T ) , (1.11a)

divu = 0 in Ω × (0, T ) . (1.11b)
Equation (1.11) together with the initial and the boundary condition
are called the incompressible Euler equations.
2. If the pressure p solely depends on the density, that is, p = p(%) (the
equation of state), then (1.3) and (1.10) together with are called the
isentropic Euler equations.
3. If the pressure p is also a function of the temperature (see Boyle’s law

for example), that is, p = p(%, ϑ) where ϑ denotes the temperature,
then an equation for ϑ is also required to complete the system.
4. Consider a very slow motion of the fluid, that is, |u| + |Du| 1 . Sup-
pose p0 and %0 are the pressure and density of the fluid in equilibrium.
Let % = %0 + %̃ , p = p0 + p̃ , where %̃ and p̃ are small. Then
1
(a) ut ≈ − Dp̃ , (b) %̃t + %0 divu ≈ 0 , (c) p̃ ≈ p 0 (%0 )%̃ .
%0
Using (c) in (b), we find that

1
(b)’ p̃t ≈ −p 0 (%0 )divu .
%0
Suppose now the fluid motion is rotation free, then u = Dϕ for some
scalar function ϕ called the velocity potential of the fluid motion. Then
(a) implies p̃ = −%0 ϕt , hence (b)’ implies
ϕtt = c2 div(Dϕ) = c2 ∆ϕ ,
p
where c2 = p 0 (%0 ) is the sound speed.
1.2.8 Minimal surface
Suppose that Ω ⊆ R2 is a bounded set with boundary parametrized by

(x(t), y(t)) for t ∈ I . Let C ⊆ R3 be a closed curve parametrized by
(x(t), y(t), f (x(t), y(t))) . We want to find a surface having C as its boundary
with minimal surface area. Then the goal is to find a function u with the
property that u = f on ∂Ω that minimizes the functional
Z p
A(w) = 1 + |Dw|2 dA .
Ω
Let ϕ ∈ C 1 (Ω) , and define
A(u + tϕ) − A(u) Du · Dϕ

Z
δA(u; ϕ) = lim = p dx .
t→0 t Ω 1 + |Du|2
If u minimize A , then δA(u; ϕ) = 0 for all ϕ ∈ Cc1 (Ω) . Assuming that

u ∈ C 2 (Ω) , we find that u satisfies
Du
div p = 0,
1 + |Du|2
or expanding the bracket using the Lebnitz rule, we obtain the minimal
surface equation
(1 + u2y )uxx − 2ux uy uxy + (1 + u2x )uyy = 0 ∀ (x, y) ∈ Ω . (1.12)

Chapter 2
Introduction to Sobolev
Spaces
2.1 Lp spaces
2.1.1 Notation
We will usually use Ω to denote an open and smooth domain in Rd , for

d = 1, 2, 3, ... In this chapter on Lp spaces, we will sometimes use X to
denote a more general measure space, but the reader can usually think of a
subset of Euclidean space.
C k (Ω) is the space of functions which are k times differentiable in Ω for

integers k ≥ 0.
C 0 (Ω) then coincides with C(Ω), the space of continuous functions on Ω.
C ∞ (Ω) = ∩k≥0 C k (Ω).
spt f denotes the support of a function f , and is the closure of the set
{x ∈ Ω | f (x) 6= 0}.
C0 (Ω) = {u ∈ C(Ω) | spt u compact in Ω}.
C0k (Ω) = C k (Ω) ∩ C0 (Ω).

C0∞ (Ω) = C ∞ (Ω)∩C0 (Ω). We will also use D(Ω) to denote this space, which
is known as the space of test functions in the theory of distributions.
11
12 CHAPTER 2. INTRODUCTION TO SOBOLEV SPACES
2.1.2 Definitions and basic properties
Definition 2.1. Let 0 < p < ∞ and let (X, M, µ) denote a measure space.
If f : X → R is a measurable function, then we define
Z 1
p
p
kf kLp (X) := |f | dx and kf kL∞ (X) := ess supx∈X |f (x)| .
X
Note that kf kLp (X) may take the value ∞. Unless stated otherwise, we will
usually consider X to be a smooth, open subset Ω of Rd , and we will assume
that all functions under consideration are measurable.
Definition 2.2. The space Lp (X) is the set
Lp (X) = {f : X → R | kf kLp (X) < ∞} .
The space Lp (X) satisfies the following vector space properties:
1. For each α ∈ R, if f ∈ Lp (X) then αf ∈ Lp (X);
2. If f, g ∈ Lp (X), then
|f + g|p ≤ 2p−1 (|f |p + |g|p ) ,
so that f + g ∈ Lp (X).
3. The triangle inequality is valid if p ≥ 1.
The most interesting cases are p = 1, 2, ∞, while all of the Lp arise often in
nonlinear estimates.
Definition 2.3. The space lp , called “little Lp ”, will be useful when we

introduce Sobolev spaces on the torus and the Fourier series. For 1 ≤ p < ∞,
we set
∞
( )
X
lp = {xn }n∈Z | |xn |p < ∞ ,
n=−∞
where Z denotes the integers.

2.1. LP SPACES 13
2.1.3 Basic inequalities
Convexity is fundamental to Lp spaces for p ∈ [1, ∞).
Lemma 2.4. For λ ∈ (0, 1), xλ ≤ (1 − λ) + λx.
Proof. Set f (x) = (1 − λ) + λx − xλ ; hence, f 0 (x) = λ − λxλ−1 = 0 if and

only if λ(1 − xλ−1 ) = 0 so that x = 1 is the critical point of f . In particular,
the minimum occurs at x = 1 with value
f (1) = 0 ≤ (1 − λ) + λx − xλ .
Lemma 2.5. For a, b ≥ 0 and λ ∈ (0, 1), aλ b1−λ ≤ λa + (1 − λ)b with

equality if a = b.
Proof. If either a = 0 or b = 0, then this is trivially true, so assume that

a, b > 0. Set x = a/b, and apply Lemma 1 to obtain the desired inequality.
Theorem 2.6 (Hölder’s inequality). Suppose that 1 ≤ p ≤ ∞ and 1 < q <

1 1
∞ with p + q = 1. If f ∈ Lp and g ∈ Lq , then f g ∈ L1 . Moreover,
kf gkL1 ≤ kf kLp kgkLq .
Note that if p = q = 2, then this is the Cauchy-Schwarz inequality since

kf gkL1 = |(f, g)L2 |.
Proof. We use Lemma 2.5. Let λ = 1/p and set

|f |p |g|q
a= , and b =
kf kpLp kgkqLp
for all x ∈ X. Then aλ b1−λ = a1/p b1−1/p = a1/p b1/q so that

|f | · |g| 1 |f |p 1 |g|q
≤ p + .
kf kLp kgkLq p kf kLp q kgkqLq
Integrating this inequality yields
|f | · |g| 1 |f |p 1 |g|q
Z
1 1
Z
dx ≤ p + q dx = + = 1 .
X kf kL kgkL
p q
X p kf kLp q kgkLq p q

p 1
Definition 2.7. The exponent q = p−1 (or q = 1− p1 ) is called the conjugate
exponent of p.
Theorem 2.8 (Minkowski’s inequality). If 1 ≤ p ≤ ∞ and f, g ∈ Lp then
kf + gkLp ≤ kf kLp + kgkLp .
Proof. If f + g = 0 a.e., then the statement is trivial. Assume that f + g 6= 0

a.e. Consider the equality
|f + g|p = |f + g| · |f + g|p−1 ≤ (|f | + |g|)|f + g|p−1 ,
and integrate over X to find that

Z Z
|f + g|p dx ≤ (|f | + |g|)|f + g|p−1 dx

X X
Hölder’s
(kf kLp + kgkLp ) |f + g|p−1 Lq .

≤
p
Since q = p−1 ,
Z 1
q
|f + g|p−1 q = p

L
|f + g| dx ,
X
from which it follows that

Z 1− 1
q
p
|f + g| dx ≤ kf kLp + kgkLq ,
X
1
which completes the proof, since p = 1 − 1q .
Corollary 2.9. For 1 ≤ p ≤ ∞, Lp (X) is a normed linear space.
Example 2.10. Let Ω denote a subset of Rn whose Lebesgue measure is

equal to one. If f ∈ L1 (Ω) satisfies f (x) ≥ M > 0 for almost all x ∈ Ω,
then log(f ) ∈ L1 (Ω) and satisfies
Z Z
log f dx ≤ log( f dx) .
Ω Ω
To see this, consider the function g(t) = t − 1 − log t for t > 0. Compute
g 0 (t) = 1− 1t = 0 so t = 1 is a minimum (since g 00 (1) > 0). Thus, log t ≤ t−1
1
and letting t 7→ t we see that
1
1− ≤ log t ≤ t − 1 . (2.1)
t
2.1. LP SPACES 15
Since log x is continuous and f is measurable, then log f is measurable for

f (x)
f > 0. Let t = kf kL1 in (2.1) to find that
kf kL1 f (x)
1− ≤ log f (x) − log kf kL1 ≤ − 1. (2.2)
f (x) kf kL1
Since g(x) ≤ log f (x) ≤ h(x) for two integrable functions g and h, it fol-
lows that log f (x) is integrable. Next, integrate (2.2) to finish the proof, as
R f (x)
X kf k 1 − 1 dx = 0.
L
2.1.4 The space (Lp (X), k · kLp (X) is complete
Recall the a normed linear space is a Banach space if every Cauchy sequence
has a limit in that space; furthermore, recall that a sequence xn → x in X
if limn→∞ kxn − xkX = 0.
The proof of completeness makes use of the following two lemmas which
are restatements of the Monotone Convergence Theorem and the Dominated
Convergence Theorem, respectively (see the Appendix for this chapter).
Lemma 2.11 (MCT). If fn ∈ L1 (X), 0 ≤ f1 (x) ≤ f2 (x) ≤ · · ·, and

kfn kL1 (X) ≤ C < ∞, then limn→∞ fn (x) = f (x) with f ∈ L1 (X) and
kfn − f kL1 → 0 as n → 0.
Lemma 2.12 (DCT). If fn ∈ L1 (X), limn→∞ fn (x) = f (x) a.e., and if

∃ g ∈ L1 (X) such that |fn (x)| ≤ |g(x)| a.e. for all n, then f ∈ L1 (X) and
kfn − f kL1 → 0.
Proof. Apply the Dominated Convergene Theorem to the sequence hn =

|fn − f | → 0 a.e., and note that |hn | ≤ 2g.
Theorem 2.13. If 1≤ p < ∞ then Lp (X) is a Banach space.
Proof. Step 1. The Cauchy sequence. Let {fn }∞

n=1 denote a Cauchy
sequence in Lp , and assume without loss of generality (by extracting a sub-
sequence if necessary) that kfn+1 − fn kLp ≤ 2−n .
Step 2. Conversion to a convergent monotone sequence. Define the

sequence {gn }∞
n=1 as
g1 = 0, gn = |f1 | + |f2 − f1 | + · · · + |fn − fn−1 | for n ≥ 2 .

It follows that
0 ≤ g1 ≤ g2 ≤ · · · ≤ gn ≤ · · ·
so that gn is a monotonically increasing sequence. Furthermore, {gn } is

uniformly bounded in Lp as
Z ∞
!p
gnp dx = kgn kpLp ≤ kf1 kLp +
X
kfi − fi−1 kLp ≤ (kf1 kLp + 1)p ;
X i=2
thus, by the Monotone Convergence Theorem, gnp % g p a.e., g ∈ Lp , and

gn ≤ g a.e.
Step 3. Pointwise convergence of {fn }. For all k ≥ 1,
|fn+k − fn | = |fn+k − fn+k−1 + fn+k−1 + · · · − fn+1 + fn+1 − fn |

n+k+1
X
≤ |fi − fi−1 | = gn+k − gn −→ 0 a.e.
i=n+1
Therefore, fn → f a.e. Since

n
X
|fn | ≤ |f1 | + |fi − fi−1 | ≤ gn ≤ g for all n ∈ N ,
i=2
it follows that |f | ≤ g a.e. Hence, |fn |p ≤ g p , |f |p ≤ g p , and |f − fn |p ≤ 2g p ,

and by the Dominated Convergence Theorem,
Z Z
p
lim |f − fn | dx = lim |f − fn |p dx = 0 .
n→∞ X X n→∞

2.1.5 Convergence criteria for Lp functions
If {fn } is a sequence in Lp (X) which converges to f in Lp (X), then there

exists a subsequence {fnk } such that fnk (x) → f (x) for almost every x ∈ X
(denoted by a.e.), but it is in general not true that the entire sequence itself
will converge pointwise a.e. to the limit f , without some further conditions
holding.
Example 2.14. Let X = [0, 1], and consider the subintervals

1 1 1 1 2 2 1 1 2 2 3 3 1
0, , , 1 , 0, , , , , 1 , 0, , , , , , , 1 , 0, , ···
2 2 3 3 3 3 4 4 4 4 4 4 5
Let fn denote the indicator function of the nth interval of the above sequence.
Then kfn kLp → 0, but fn (x) does not converge for any x ∈ [0, 1].
2.1. LP SPACES 17
Example 2.15. Set X = R, and for n ∈ N, set fn = 1[n,n+1] . Then

fn (x) → 0 as n → ∞, but kfn kLp = 1 for p ∈ [1, ∞); thus, fn → 0
pointwise, but not in Lp .
Example 2.16. Set X = [0, 1], and for n ∈ N, set fn = n1[0, 1 ] . Then
n
fn (x) → 0 a.e. as n → ∞, but kfn kL1 = 1; thus, fn → 0 pointwise, but not
in L1 .
Theorem 2.17. For 1 ≤ p < ∞, suppose that {fn } ⊂ Lp (X) and that
fn (x) → f (x) a.e. If limn→∞ kfn kLp (X) = kf kLp (X) , then fn → f in Lp (X).
a+b p
≤ 12 (ap + bp ) so that

Proof. Given a, b ≥ 0, convexity implies that 2
(a + b)p ≤ 2p−1 (ap + bp ), and hence |a − b|p ≤ 2p−1 (|a|p + |b|p ). Set a = fn
and b = f to obtain the inequality
0 ≤ 2p−1 (|fn |p + |f |p ) − |fn − f |p .
Since fn (x) → f (x) a.e.,

Z Z
p p
lim 2p−1 (|fn |p + |f |p ) − |fn − f |p dx .

2 |f | dx =
X X n→∞
Thus, Fatou’s lemma asserts that

Z Z
p p
2p−1 (|fn |p + |f |p ) − |fn − f |p dx

2 |f | dx ≤ lim inf
X n→∞
Z X Z Z
p−1 p p−1 p p
=2 |f | dx + 2 lim |fn | + lim inf − |fn − f | dx
X n→∞ X n→∞ X
Z Z
p−1 p
=2 |f | dx − lim sup |fn − f |p dx .
X n→∞ X
|f |p dx < ∞, the last inequality shows that lim supn→∞ X |fn −

R R
As X
f |p dx ≤ 0. It follows that lim supn→∞ X |fn − f |p dx = lim inf n→∞ X |fn −
R R
f |p dx = 0, so that limn→∞ X |fn − f |p dx = 0.

R

2.1.6 The space L∞ (X)
Definition 2.18. With kf kL∞ (X) = inf{M ≥ 0 | |f (x)| ≤ M a.e.}, we set
L∞ (X) = {f : X → R | kf kL∞ (X) < ∞} .
Theorem 2.19. (L∞ (X), k · kL∞ (X) ) is a Banach space.

Proof. Let fn be a Cauchy sequence in L∞ (X). It follows that |fn − fm | ≤

kfn − fm kL∞ (X) a.e. and hence fn (x) → f (x) a.e., where f is measurable
and essentially bounded.
Choose > 0 and N () such that kfn −fm kL∞ (X) < for all n, m ≥ N ().
Since |f (x) − fn (x)| = limm→∞ |fm (x) − fn (x)| ≤ holds a.e. x ∈ X, it
follows that kf − fn kL∞ (X) ≤ for n ≥ N (), so that kfn − f kL∞ (X) → 0.
Remark 2.20. In general, there is no relation of the type Lp ⊂ Lq . For

1
example, suppose that X = (0, 1) and set f (x) = x− 2 . Then f ∈ L1 (0, 1),
but f 6∈ L2 (0, 1). On the other hand, if X = (1, ∞) and f (x) = x−1 , then
f ∈ L2 (1, ∞), but f 6∈ L1 (1, ∞).
Lemma 2.21 (Lp comparisons). If 1 ≤ p < q < r ≤ ∞, then (a) Lp ∩ Lr ⊂

Lq , and (b) Lq ⊂ Lp + Lr .
Proof. We begin with (b). Suppose that f ∈ Lq , define the set E = {x ∈

X : |f (x)| ≥ 1}, and write f as
f = f 1E + f 1E c
= g + h.
Our goal is to show that g ∈ Lp and h ∈ Lr . Since |g|p = |f |p 1E ≤ |f |q 1E

and |h|r = |f |r 1E c ≤ |f |q 1E c , assertion (b) is proven.
For (a), let λ ∈ [0, 1] and for f ∈ Lq ,
Z 1 Z 1
q q
q λq (1−λ)q
kf kLq = |f | dx = |f | |f | dµ
X X
1
(1−λ)q (1−λ)
kf kλq
q
≤ Lp kf kLr = kf kλLp kf kLr .
Theorem 2.22. If µ(X) ≤ ∞ and q > p, then Lq ⊂ Lp .
Proof. Consider the case that q = 2 and p = 1. Then by the Cauchy-Schwarz

inequality, Z Z p
|f |dx = |f | · 1 dx ≤ kf kL2 (X) µ(X) .
X X

2.1. LP SPACES 19
2.1.7 Approximation of Lp (X) by simple functions

Pn
Lemma 2.23. If p ∈ [1, ∞), then the set of simple functions f = i=1 ai 1Ei ,
where each Ei is an element of the σ-algebra A and µ(Ei ) < ∞, is dense in
Lp (X, A, µ).
Proof. If f ∈ Lp , then f is measurable; thus, there exists a sequence {φn }∞

n=1
of simple functions, such that φn → f a.e. with
0 ≤ |φ1 | ≤ |φ2 | ≤ · · · ≤ |f |,
i.e., φn approximates f from below.

Recall that |φn − f |p → 0 a.e. and |φn − f |p ≤ 2p |f |p ∈ L1 , so by the
Dominated Convergence Theorem, kφn − f kLp → 0.
Now, suppose that the set Ei are disjoint; then by the definition of the
Lebesgue integral,
Z n
X
φpn dx = |ai |p µ(Ei ) < ∞ .
X i=1
If ai 6= 0, then µ(Ei ) < ∞.
2.1.8 Approximation of Lp (Ω) by continuous functions
Lemma 2.24. Suppose that Ω ⊂ Rn is bounded. Then C 0 (Ω) is dense in

Lp (Ω) for p ∈ [1, ∞).
Proof. Let K be any compact subset of Ω. The functions
1
FK ,n (x) = ∈ C 0 (Ω) satisfy FK ,n ≤ 1 ,
1 + n dist(x, K)
and decrease monotonically to the characteristic function 1K . The Mono-
tone Convergence Theorem gives
fK ,n → 1K in Lp (Ω), 1 ≤ p < ∞ .
Next, let A ⊂ Ω be any measurable set, and let λ denote the Lebesgue
measure. Then
λ(A) = sup{µ(K) : K ⊂ A, K compact} .

It follows that there exists an increasing sequence of Kj of compact subsets

of A such that λ(A\ ∪j Kj ) = 0. By the Monotone Convergence Theorem,
1Kj → 1A in Lp (Ω) for p ∈ [1, ∞). According to Lemma 2.23, each function
in Lp (Ω) is a norm limit of simple functions, so the lemma is proved.
2.1.9 Approximation of Lp (Ω) by smooth functions
For Ω ⊂ Rn open, for > 0 taken sufficiently small, define the open subset
of Ω by
Ω := {x ∈ Ω | dist(x, ∂Ω) > } .
Definition 2.25 (Mollifiers). Define η ∈ C ∞ (Rn ) by

2 −1
Ce(|x| −1)

if |x| < 1
η(x) := ,
0 if |x| ≥ 1
R
with constant C > 0 chosen such that Rn η(x)dx = 1.
For > 0, the standard sequence of mollifiers on Rn is defined by
η (x) = −n η(x/) ,

R
and satisfy Rn η (x)dx = 1 and spt(η ) ⊂ B(0, ).
Definition 2.26. For Ω ⊂ Rn open, set
Lploc (Ω) = {u : Ω → R | u ∈ Lp (Ω̃) ∀ Ω̃ ⊂⊂ Ω} ,
where Ω̃ ⊂⊂ Ω means that there exists K compact such that Ω̃ ⊂ K ⊂ Ω.

We say that Ω̃ is compactly contained in Ω.
Definition 2.27 (Mollification of L1 ). If f ∈ L1loc (Ω), define its mollifica-

tion
f = η ∗ f in Ω ,
so that
Z Z
f (x) = η (x − y)f (y)dy = η (y)f (x − y)dy ∀x ∈ Ω .
Ω B(0,)
Theorem 2.28 (Mollification of Lp (Ω)).
(A) f ∈ C ∞ (Ω ).
2.1. LP SPACES 21
(B) f → f a.e. as → 0.
(C) If f ∈ C 0 (Ω), then f → f uniformly on compact subsets of Ω.
(D) If p ∈ [1, ∞) and f ∈ Lploc (Ω), then f → f in Lploc (Ω).
Proof. Part (A). We rely on the difference quotient approximation of the

partial derivative. Fix x ∈ Ω , and choose h sufficiently small so that x +
hei ∈ Ω for i = 1, ..., n, and compute the difference quotient of f :
f (x + hei ) − f (x) x + hei − y x−y

1
Z
= −n η −η f (y)dy
h Ω h
x + hei − y x−y

1
Z
−n
= η −η f (y)dy
Ω̃ h
for some open set Ω̃ ⊂⊂ Ω. On Ω̃,
x + hei − y x−y 1 ∂η x − y

1
lim η −η = ,
h→0 h ∂xi
so by the Dominated Convergence Theorem,
∂f ∂η
Z
(x) = (x − y)f (y)dy .
∂xi Ω ∂xi
A similar argument for higher-order partial derivatives proves (A).
Step 2. Part (B). By the Lebesgue differentiation theorem,
1
Z
lim |f (y) − f (x)|dy = 0 for a.e. x ∈ Ω .
→0 |B(x, )| B(x,)
Choose x ∈ Ω for which this limit holds. Then

Z
|f (x) − f (x)| ≤ η (x − y)|f (y) − f (x)|dy
B(x,)
1
Z
= n η((x − y)/)|f (y) − f (x)|dy
B(x,)
C
Z
≤ |f (x) − f (y)|dy −→ 0 as → 0.
|B(x, )| B(x,)
Step 3. Part (C). For Ω̃ ⊂ Ω, the above inequality shows that if f ∈ C 0 (Ω)
and hence uniformly continuous on Ω̃, then f (x) → f (x) uniformly on Ω̃.
Step 4. Part (D). For f ∈ Lploc (Ω), p ∈ [1, ∞), choose open sets U ⊂⊂
D ⊂⊂ Ω; then, for > 0 small enough,
kf kLp (U ) ≤ kf kLp (D) .
To see this, note that

Z
|f (x)| ≤ η (x − y)|f (y)|dy
B(x,)
Z
= η (x − y)(p−1)/p η (x − y)1/p |f (y)|dy
B(x,)
Z !(p−1)/p Z !1/p
p
≤ η (x − y)dy η (x − y)|f (y)| dy ,
B(x,) B(x,)
so that for > 0 sufficiently small

Z Z Z
p
|f (x)| dx ≤ η (x − y)|f (y)|p dydx
U U B(x,)
Z Z ! Z
p
≤ |f (y)| η (x − y)dx dy ≤ |f (y)|p dy .
D B(y,) D
Since C 0 (D) is dense in Lp (D), choose g ∈ C 0 (D) such that kf −

gkLp (D) < δ; thus
kf − f kLp (U ) ≤ kf − g kLp (U ) + kg − gkLp (U ) + kg − f kLp (U )

≤ 2kf − gkLp (D) + kg − gkLp (U ) ≤ 2δ + kg − gkLp (U ) .
2.1.10 Continuous linear functionals on Lp (X)
Let Lp (X)0 denote the dual space of Lp (X). For φ ∈ Lp (X)0 , the operator
norm of φ is defined by kφkop = supLp (X)=1 |φ(f )|.
p
Theorem 2.29. Let p ∈ (1, ∞], q = p−1 . For g ∈ Lq (X), define Fg :
Lp (X) → R as Z
Fg (f ) = f gdx .
X
Then Fg is a continuous linear functional on Lp (X) with operator norm
kFg kop = kgkLq (X) .
2.1. LP SPACES 23
Proof. The linearity of Fg again follows from the linearity of the Lebesgue
integral. Since
Z Z

|Fg (f )| = f gdx ≤ |f g| dx ≤ kf kLp kgkLq ,
X X
with the last inequality following from Hölder’s inequality, we have that
supkf kLp =1 |Fg (f )| ≤ kgkLq .
For the reverse inequality let f = |g|q−1 sgn g. f is measurable and in Lp
q
since |f |p = |f | q−1 = |g|q and since f g = |g|q ,
Z Z Z 1+1
p q
q q
Fg (f ) = f gdx = |g| dx = |g| dx
X X X
Z 1 Z 1
p q
= |f |p dx |g q | dx = kf kLp kgkLq
X X
Fg (f )
so that kgkLq = ≤ kFg kop .
kf kLp

Remark 2.30. Theorem 2.29 shows that for 1 < p ≤ ∞, there exists a
linear isometry g 7→ Fg from Lq (X) into Lp (X)0 , the dual space of Lp (X).
When p = ∞, g 7→ Fg : L1 (X) → L∞ (X)0 is rarely onto (L∞ (X)0 is strictly
larger than L1 (X)); on the other hand, if the measure space X is σ-finite,
then L∞ (X) = L1 (X)0 .
2.1.11 A theorem of F. Riesz
Theorem 2.31 (Representation theorem). Suppose that 1 < p < ∞ and

p
φ ∈ Lp (X)0 . Then there exists g ∈ Lq (X), q = p−1 such that
Z
φ (f ) = f gdx ∀f ∈ Lp (X) ,
X
and kφkop = kgkLq .
Corollary 2.32. For p ∈ (1, ∞) the space Lp (X, µ) is reflexive, i.e., Lp (X)00 =
Lp (X).
The proof Theorem 2.31 crucially relies on the Radon-Nikodym theorem,

whose statement requires the following definition.
Definition 2.33. If µ and ν are measure on (X, A) then ν µ if ν(E) = 0

for every set E for which µ(E) = 0. In this case, we say that ν is absolutely
continuous with respect to µ.
Theorem 2.34 (Radon-Nikodym). If µ and ν are two finite measures on

X, i.e., µ(X) < ∞, ν(X) < ∞, and ν µ, then
Z Z
F (x) dν(x) = F (x)h(x)dµ(x) (2.3)
X X
holds for some nonnegative function h ∈ L1 (X, µ) and every positive mea-
surable function F .
Proof. Define measures α = µ+2ν and ω = 2µ+ν, and let H = L2 (X,

Z α) (a
Hilbert space) and suppose φ : L2 (X, α) → R is defined by φ (f ) = f dω.
X
We show that φ is a bounded linear functional since
Z Z Z

|φ(f )| = f d(2µ + ν) ≤
|f | d(2µ + 4ν) = 2 |f | dα
X X X
p
≤ kf kL2 (x,α) α(X) .
Thus, by the Riesz representation theorem, there exists g ∈ L2 (X, α)

such that Z Z
φ(f ) = f dω = f g dα ,
X X
which implies that
Z Z
f (2g − 1)dν = f (2 − g)dµ . (2.4)
X X
F
Given 0 ≤ F a measurable function on X, if we set f = 2g−1 and
2−g R R
h = 2g−1 then X F dν = X F h dx which is the desired result, if we can
prove that 1/2 ≤ g(x) ≤ 2. Define the sets

1 1 1 2 1
En = x ∈ X | g(x) < − and En = x ∈ X | g(x) > 2 + .
2 n n
By substituting f = 1Enj , j = 1, 2 in (2.4), we see that
µ(Enj ) = ν(Enj ) = 0 for j = 1, 2 ,
from which the bounds 1/2 ≤ g(x) ≤ 2 hold. Also µ({x ∈ X | g(x) =
1/2}) = 0 and ν({x ∈ X | g(x) = 2}) = 0. Notice that if F = 1, then
h ∈ L1 (X).
2.1. LP SPACES 25
Remark 2.35. The more general version of the Radon-Nikodym theorem.

Suppose that µ(X) < ∞, ν is a finite signed measure (by the Hahn decom-
position, ν = ν − + ν + ) such that ν µ; then, there exists h ∈ L1 (X, µ)
R R
such that X F dν = X F h dµ.
Lemma 2.36 (Converse to Hölder’s inequality). Let µ(X) < ∞. Suppose

that g is measurable and f g ∈ L1 (X) for all simple functions f . If
Z

M (g) = sup f g dµ : f is a simple function < ∞ , (2.5)

kf kLp =1 X
then g ∈ Lq (X), and kgkLq (X) = M (g).
Proof. Let φn be a sequence of simple functions such that φn → g a.e. and

|φn | ≤ |g|. Set
|φn |q−1 sgn (φn )
fn =
kφn kq−1
Lq
so that kfn kLp = 1 for p = q/(q − 1). By Fatou’s lemma,
Z
kgkLq (X) ≤ lim inf kφn kLq (X) = lim inf |fn φn |dµ .
n→∞ n→∞ X
Since φn → g a.e., then
Z Z
kgkLq (X) ≤ lim inf |fn φn |dµ ≤ lim inf |fn g|dµ ≤ M (g) .
n→∞ X n→∞ X
The reverse inequality is implied by Hölder’s inequality.
Proof of the Lp (X)0 representation theorem. We have already proven that

there exists a natural inclusion ι : Lq (X) → Lp (X)0 which is an isometry. It
remains to show that ι is surjective.
Let φ ∈ Lp (X)0 and define a set function ν on measurable subsets E ⊂ X
by Z
ν(E) = 1E dν =: φ(1E ) .
X
Thus, if µ(E) = 0, then ν(E) = 0. Then
Z
f dν =: φ(f )
X
for all simple functions f , and by Lemma 2.23, this holds for all f ∈ Lp (X).
By the Radon-Nikodym theorem, there exists 0 ≤ g ∈ L1 (X) such that
Z Z
f dν = f g dµ ∀ f ∈ Lp (X) .
X X
But Z Z
φ(f ) = f dν = f g dµ (2.6)
X X
and since φ ∈ Lp (X)0 , then M (g) given by (2.5) is finite, and by the converse
to Hölder’s inequality, g ∈ Lq (X), and kφkop = M (g) = kgkLq (X) .
2.1.12 Weak convergence
The importance of the Representation Theorem 2.31 is in the use of the

weak-* topology on the dual space Lp (X)0 . Recall that for a Banach space
∗
B and for any sequence φj in the dual space B0 , φj * φ in B0 weak-*, if
hφj , f i → hφ, f i for each f ∈ B, where h·, ·i denotes the duality pairing
between B0 and B.
Theorem 2.37 (Alaoglu’s Lemma). If B is a Banach space, then the closed

unit ball in B0 is compact in the weak -* topology.
Definition 2.38. For 1 ≤ p < ∞, a sequence {fn } ⊂ Lp (X) is said to

weakly converge to f ∈ Lp (X) if
p
Z Z
fn (x)φ(x)dx → f (x)φ(x)dx ∀φ ∈ Lq (X), q = .
X X p−1
We denote this convergence by saying that fn * f in Lp (X) weakly.
Given that Lp (X) is reflexive for p ∈ (1, ∞), a simple corollary of

Alaoglu’s Lemma is the following
Theorem 2.39 (Weak compactness for Lp , 1 < p < ∞). If 1 < p < ∞ and
{fn } is a bounded sequence in Lp (X), then there exists a subsequence {fn k }
such that fn k * f in Lp (X) weakly.
Definition 2.40. A sequence {fn } ⊂ L∞ (X) is said to converge weak-* to

f ∈ L∞ (X) if
Z Z
fn (x)φ(x)dx → f (x)φ(x)dx ∀φ ∈ L1 (X) .
X X
∗
We denote this convergence by saying that fn * f in L∞ (X) weak-*.
Theorem 2.41 (Weak-* compactness for L∞ ). If {fn } is a bounded se-

∗
quence in L∞ (X), then there exists a subsequence {fn k } such that fn k * f
in L∞ (X) weak-*.
2.1. LP SPACES 27
Lemma 2.42. If fn → f in Lp (X), then fn * f in Lp (X).
Proof. By Hölder’s inequality,

Z

g(fn − f )dx ≤ kfn − f kLp kgkLq .

X
Note that if fn is weakly convergent, in general, this does not imply that
fn is strongly convergent.
Example 2.43. If p = 2, let fn denote any orthonormal sequence in L2 (X).

From Bessel’s inequality
∞ Z
X

fn gdx ≤ kgk2 2

L (X) ,
n=1 X
we see that fn * 0 in L2 (X).
This example shows that the map f 7→ kf kLp is continuous, but not
weakly continuous. It is, however, weakly lower-semicontinuous.
Theorem 2.44. If fn * f weakly in Lp (X), then kf kLp ≤ lim inf n→∞ kfn kLp .
Proof. As a consequence of Theorem 2.31,

Z Z

kf kLp (X) = sup f gdx = sup lim fn gdx
kgkLq (X) =1 X kgkLq (X) =1 n→∞ X
≤ sup lim inf kfn kLp kgkLq .

kgkLq (X) =1 n→∞
Theorem 2.45. If fn * f in Lp (X), then fn is bounded in Lp (X).
Theorem 2.46. Suppose that Ω ⊂ Rn is a bounded. Suppose that
sup kfn kLp (Ω) ≤ M < ∞ and fn → f a.e.

n
If 1 < p < ∞, then fn * f in Lp (Ω).

Proof. Egoroff’s theorem states that for all > 0, there exists E ⊂ Ω such
that µ(E) < and fn → f uniformly on E c . By definition, fn * f in Lp (Ω)
p
for p ∈ (1, ∞) if Ω (fn − f )gdx → 0 for all g ∈ Lq (Ω), q = p−1
R
. We have the
inequality
Z Z Z
(fn − f )gdx ≤ |fn − f | |g| dx + |fn − f | |g| dx .
Ω E Ec
Choose n ∈ N sufficiently large, so that |fn (x) − f (x)| ≤ δ for all x ∈ E c .

By Hölder’s inequality,
Z
|fn − f | |g| dx ≤ kfn − f kLp (E c ) kgkLq (E c ) ≤ δµ(E c )kgkLq (Ω) ≤ Cδ
Ec
for a constant C < ∞.

By the Dominated Convergence Theorem, kfn − f kLp (Ω) ≤ 2M so by
Hölder’s inequality, the integral over E is bounded by 2M kgkLq (E) . Next, we
use the fact that the integral is continuous with respect to the measure of the
set over which the integral is taken. In particular, if 0 ≤ h is integrable, then
for all δ > 0, there exists > 0 such that if the set E has measure µ(E ) < ,
R
then E hdx ≤ δ. To see this, either approximate h by simple functions, or
R
use the Dominated Convergence theorem for the integral Ω 1E (x)h(x)dx.
Remark 2.47. The proof of Theorem 2.46 does not work in the case that
p = 1, as Hölder’s inequality gives
Z
|fn − f | |g| dx ≤ kfn − f kL1 (Ω) kgkL∞ (E) ,
E
so we lose the smallness of the right-hand side.
Remark 2.48. Suppose that E ⊂ X is bounded and measurable, and let

g = 1E . If fn * f in Lp (X), then
Z Z
fn (x)dx → f (x)dx;
E E
hence, if fn * f , then the average of fn converges to the average of f

pointwise.
2.1. LP SPACES 29
2.1.13 Integral operators
If u : Rn → R satisfies certain integrability conditions, then we can define

the operator K acting on the function u as follows:
Z
Ku(x) = k(x, y)u(y)dy ,
Rn
where k(x, y) is called the integral kernel.The mollification procedure, intro-
duced in Definition 2.27, is one example of the use of integral operators; the
Fourier transform is another.
Definition 2.49. Let L(Lp (Rn ), Lp (Rn )) denote the space of bounded linear
operators from Lp (Rn ) to itself. Using the Representation Theorem 2.31, the
natural norm on L(Lp (Rn ), Lp (Rn )) is given by
Z

kKkL(Lp (Rn ),Lp (Rn )) = sup sup Kf (x)g(x)dx .
kf kLp =1 kgkLq =1 n
R
R
Theorem 2.50. Let 1 ≤ p < ∞, Ku(x) = Rn k(x, y)u(y)dy, and suppose
that
Z Z
n
|k(x, y)|dx ≤ C1 ∀y ∈ R and |k(x, y)|dy ≤ C2 ∀x ∈ Rn ,
Rn Rn
where 0 < C1 , C2 < ∞. Then K : Lp (Rn ) → Lp (Rn ) is bounded and
1 p−1
kKkL(Lp (Rn ),Lp (Rn )) ≤ C1p C2 p .
In order to prove Theorem 2.50, we will need another well-known in-

equality.
Lemma 2.51 (Cauchy-Young Inequality). If p1 + 1q = 1, then for all a, b ≥ 0,

ap bq
ab ≤ + .
p q
Proof. Suppose that a, b > 0, otherwise the inequality trivially holds.
ab = exp(log(ab)) = exp(log a + log b) (since a, b > 0)

1 p 1 q
= exp log a + log b
p q
1 1
≤ exp(log ap ) + exp(log bq ) (using the convexity of exp)
p q
ap bq
= +
p q
1 1
where we have used the condition p + q = 1.
1
Lemma 2.52 (Cauchy-Young Inequality with δ). If p + 1q = 1, then for all
a, b ≥ 0,
ab ≤ δ ap + Cδ bq , δ > 0,
with Cδ = (δp)−q/p q −1 .
Proof. This is a trivial consequence of Lemma 2.51 by setting

b
ab = a · (δp)1/p .
(δp)1/p

|f (y)|p
Proof of Theorem 2.50. According to Lemma 2.51, |f (y)g(x)| ≤ p +
|g(x)|q
q so that
Z Z

n n k(x, y)f (y)g(x)dydx

R R
|k(x, y)| |k(x, y)|
Z Z Z Z
p
≤ dx|f (y)| dy + dy|g(x)|q dx
Rn Rn p Rn Rn q
C1 C2
≤ kf kpLp + kgkqLq .
p q
To improve this bound, notice that
Z Z

n n k(x, y)f (y)g(x)dydx

R R
|k(x, y)| |k(x, y)|
Z Z Z Z
≤ p
dx|tf (y)| dy + dy|t−1 g(x)|q dx
Rn Rn p Rn Rn q
C1 tp C2 t−q
≤ kf kpLp + kgkqLq =: F (t) .
p q
Find the value of t for which F (t) has a minimum to establish the desired
bounded.
Theorem 2.53 (Simple version of Young’s inequality). Suppose that k ∈

L1 (Rn ) and f ∈ Lp (Rn ). Then
kk ∗ f kLp ≤ kkkL1 kf kLp .
Proof. Define Z
Kk (f ) = k ∗ f := k(x − y)f (y)dy .
Rn
Let C1 = C2 = kkkL1 (Rn ) . Then according to Theorem 2.50, Kk : Lp (Rn ) →
Lp (Rn ) and kKk kL(Lp (Rn ),Lp (Rn )) ≤ C1 .
2.1. LP SPACES 31
Theorem 2.50 can easily be generalized to the setting of integral op-

erators K : Lq (Rn ) → Lr (Rn ) built with kernels k ∈ Lp (Rn ) such that
1 1
1+ r = p + 1q . Such a generalization leads to
Theorem 2.54 (Young’s inequality). Suppose that k ∈ Lp (Rn ) and f ∈

Lq (Rn ). Then
1 1 1
kk ∗ f kLr ≤ kkkLp kf kLq for 1 + = + .
r p q
2.1.14 Appendix 1: The monotone and dominated conver-

gence theorems and Fatou’s lemma
Let Ω ⊂ Rd denote an open and smooth subset. The domain Ω is called

smooth whenever its boundary ∂Ω is a smooth (d − 1)-dimensional hyper-
surface.
Theorem 2.55 (Monotone Convergence Theorem). Let fn : Ω → R∪{+∞}

denote a sequence of functions, fn ≥ 0, and suppose that the sequence fn is
monotonically increasing, i.e.,
f1 ≤ f2 ≤ f3 ≤ · · ·
Then Z Z
lim fn (x)dx = lim fn (x)dx .
n→∞ Ω Ω n→∞
Lemma 2.56 (Fatou’s Lemma). Suppose the sequence fn : Ω → R and

fn ≥ 0. Then
Z Z
lim inf fn (x)dx ≤ lim inf fn (x)dx .
Ω n→∞ n→∞ Ω
Example 2.57. Consider Ω = (0, 1) ⊂ R and suppose that fn = n1(0,1/n) .

R1 R1
Then 0 fn (x)dx = 1 for all n ∈ N, but lim inf n→∞ 0 fn (x)dx = 0.
Theorem 2.58 (Dominated Convergence Theorem). Suppose the sequence

fn : Ω → R, fn (x) → f (x) almost everywhere (with respect to Lebesgue
measure), and furthermore, |fn | ≤ g ∈ L1 (Ω). Then f ∈ L1 (Ω) and
Z Z
lim fn (x)dx = f (x)dx .
n→∞ Ω Ω
Equivalently, fn → f in L1 (Ω) so that limn→∞ kfn − f kL1 (Ω) = 0.

In the exercises, you will be asked to prove that the Monotone Conver-
gence Theorem implies Fatou’s Lemma which, in turn, implies the Domi-
nated Convergence Theorem.
2.1.15 Appendix 2: The Fubini and Tonelli Theorems
Let (X, A, µ) and (Y, B, ν) denote two fixed measure spaces. The product
σ-algebra A × B of subsets of X × Y is defined by
A × B = {A × B : A ∈ A, B ∈ B}.
The set function µ × ν : A × B → [0, ∞] defined by
(µ × ν)(A × B) = µ(A) · ν(B)
for each A × B ∈ A × B is a measure.
Theorem 2.59 (Fubini). Let f : X × Y → R be a µ × ν-integrable function.

Then both iterated integrals exist and
Z Z Z Z Z
f d(µ × ν) = f dµdν = f dνdµ .
X×Y Y X X Y
The existence of the iterated integrals is by no means enough to ensure

that the function is integrable over the product space. As an example, let
X = Y = [0, 1] and µ = ν = λ with λ the Lebesgue measure. Set
(
x2 −y 2
(x2 +y 2 )2
,(x, y) 6= (0, 0)
f (x, y) = .
0, (x, y) = (0, 0)
Then a standard computation shows that

1Z 1 1Z 1
π π
Z Z
f (x, y)dxdy = − , f (x, y)dydx = .
0 0 4 0 0 4
Fubini’s theorem shows, of course, that f is not integrable over [0, 1]2
There is a converse to Fubini’s theorem, however, according to which
the existence of one of the iterated integrals is sufficient for the integrability
of the function over the product space. The theorem is known as Tonelli’s
theorem, and this result is often used.
2.1. LP SPACES 33
Theorem 2.60 (Tonelli). Let (X, A, µ) and (Y, B, ν) denote two σ-finite
measure spaces, and let f : X × Y → R be a µ × ν-measurable function. If
R R R R
one of the iterated integrals X Y |f |dνdµ or Y X |f |dµdν exists, then the
function f is µ × ν-integrable and hence, the other iterated integral exists
and Z Z Z Z Z
f d(µ × ν) = f dµdν = f dνdµ .
X×Y Y X X Y
2.1.16 Exercises
Problem 2.1. Use the Monotone Convergence Theorem to prove Fatou’s

Lemma.
Problem 2.2. Use Fatou’s Lemma to prove the Dominated Convergence

Theorem.
Problem 2.3. Let Ω ⊂ Rd denote an open and smooth subset. Let (a, b) ⊂ R
be an open interval, and let f : (a, b) × Ω → R be a function such that for
df
each t ∈ (a, b), f (t, ·) : Ω → R is integrable and dt (t, x) exists for each
(t, x) ∈ (a, b) × Ω. Futhermore, assume that there is an integrable function
g : Ω → [0, ∞) such that supt∈(a,b) | df (t, x)| ≤ g(x) for all x ∈ Ω. Show that
R dt
the function h defined by h(t) ≡ Ω f (t, x)dx is differentiable and that the
derivative is given by
dh d df
Z Z
(t) = f (t, x)dx = (t, x)dx
dt dt Ω Ω dt
for each t ∈ (a, b). Hint: You will need to use the definition of the deriva-
dr
tive for a real valued function function r : (a, b) → R which is dt (t0 ) =
limh→0 r(t0 +h)−r(t
h
0)
, as well as the Mean Value Theorem from calculus which
states the following: Let (t1 , t2 ) ⊂ R and let q : (t1 , t2 ) → R be differentiable
|q(t2 )−q(t1 )| dq 0
on (t1 , t2 ). Then t2 −t1 = dt (t ) where t1 is some point between t1 and
t2 .
Problem 2.4. Let Ω denote an open subset of Rn . If f ∈ L1 (Ω) ∩ L∞ (Ω),

show that f ∈ Lp (Ω) for 1 < p < ∞. If Ω is bounded, then show that
limp%∞ kf kLp = kf kL∞ . (Hint: For > 0, you can prove that the set
E = {x ∈ Ω : |f (x)| > kf kL∞ − } has positive Lebesgue measure, and the
inequality [kf kL∞ − ] 1E ≤ |f | holds.)
Problem 2.5. Theorem 2.17 states that if 1 ≤ p < ∞, f ∈ Lp , {fn } ⊂ Lp ,

fn → f a.e., and limn→∞ kfn kLp = kf kLp , then limn→∞ kfn − f kLp → 0.
Show by an example that this theorem is false when p = ∞.
Problem 2.6. Show that equality holds in the inequality
aλ b1−λ ≤ λa + (1 − λ)b, λ ∈ (0, 1), a, b ≥ 0
if and only if a = b. Use this to show that if f ∈ Lp and g ∈ Lq for

1 1
1 < p, q < ∞ and p + q = 1, then
Z
|f g|dx = kf kLp kgkLq
Ω
holds if and only if there exists two constants C1 and C2 (not both zero) such
that C1 |f |p = C2 |g|q holds.
Problem 2.7. Use the result of Problem 2.6 to prove that if f, g ∈ L3 (Ω)
satisfy Z
kf kL3 = kgkL3 = f 2 g dx = 1 ,
Ω
then g = |f | a.e.
Problem 2.8. Given f ∈ L1 (S1 ), 0 < r < 1, define

∞ 2π
1
Z
fˆn r|n| einθ , fˆn =
X
Pr f (θ) = f (θ)e−inθ dθ .
n=−∞
2π 0
Show that
2π
1
Z
Pr f (θ) = pr ∗ f (θ) = pr (θ − φ)f (φ)dφ ,
2π 0
where
∞
X 1 − r2
pr (θ) = r|n| einθ = .
n=−∞
1 − 2r cos θ + r2
1
R 2π
Show that 2π 0 pr (θ)dθ = 1.
2.1. LP SPACES 35
Problem 2.9. If f ∈ Lp (S1 ), 1 ≤ p < ∞, show that
Pr f → f in Lp (S1 ) as r % 1 .
Problem 2.10. Suppose that Y = [0, 1]2 is the unit square in R2 and let
a(y) denote a Y -periodic function in L∞ (R2 ). For > 0, let a (x) = a( x ),
R ∗
and let ā = Y a(y)dy denote the average value of a. Prove that a * ā as
→ 0.
√
Problem 2.11. Let fn = n1(0, 1 ) . Prove that fn * 0 in L2 (0, 1), that
n
fn → 0 in L1 (0, 1), but that fn does not converge strongly in L2 (0, 1).
2.2 The Sobolev spaces H k (Ω) for integers k ≥ 0

2.2.1 Weak derivatives
Definition 2.61 (Test functions). For Ω ⊂ Rn , set
C0∞ (Ω) = {u ∈ C ∞ (Ω) | spt(u) ⊂ V ⊂⊂ Ω},
the smooth functions with compact support. Traditionally D(Ω) is often used
to denote C0∞ (Ω), and D(Ω) is often referred to as the space of test functions.
du
For u ∈ C 1 (R), we can define dx by the integration-by-parts formula;
namely,
du dφ
Z Z
(x)φ(x)dx = − u(x) (x)dx ∀φ ∈ C0∞ (R) .
R dx R dx
Notice, however, that the right-hand side is well-defined, whenever u ∈

L1loc (R)
Definition 2.62. An element α ∈ Zn+ (nonnegative integers) is called a

∂ α1 ∂ αn
multi-index. For such an α = (α1 , ..., αn ), we write Dα = α
∂x1 1
··· ∂xα n and
n
|α| = α1 + · · ·αn .
Example 2.63. Let n = 2. If |α| = 0, then α = (0, 0); if |α| = 1, then

α = (1, 0) or α = (0, 1). If |α| = 2, then α = (1, 1).
Definition 2.64 (Weak derivative). Suppose that u ∈ L1loc (Ω). Then v α ∈

L1loc (Ω) is called the αth weak derivative of u, written v α = Dα u, if
Z Z
|α|
α
u(x) D φ(x)dx = (−1) v α (x)φ(x)dx ∀φ ∈ C0∞ (Ω) .
Ω Ω
Example 2.65. Let n = 1 and set Ω = (0, 2). Define the function

x, 0 ≤ x < 1
u(x) = .
1, 1 ≤ x ≤ 2
Then the function

1, 0 ≤ x < 1
v(x) =
0, 1 ≤ x ≤ 2
2.2. THE SOBOLEV SPACES H K (Ω) FOR INTEGERS K ≥ 0 37
is the weak derivative of u. To see this, note that for φ ∈ C0∞ (0, 2),
2 1 2
dφ dφ dφ
Z Z Z
u(x) (x)dx = x (x)dx + (x)dx
0 dx 0 dx 1 dx
Z 1 Z 1
=− φ(x)dx + xφ|10 + φ|21 =− φ(x)dx
0 0
Z 2
=− v(x)φ(x)dx .
0
Example 2.66. Let n = 1 and set Ω = (0, 2). Define the function

x, 0 ≤ x < 1
u(x) = .
2, 1 ≤ x ≤ 2
Then the weak derivative does not exist!

To prove this, assume for the sake of contradiction that there exists v ∈
L1loc (Ω) such that for all φ ∈ C0∞ (0, 2),
2 2
dφ
Z Z
v(x)φ(x)dx = − u(x) (x)dx .
0 0 dx
Then
2 Z 1 Z 2
dφ dφ
Z
v(x)φ(x)dx = − x (x)dx − 2 (x)dx
0 0 dx 1 dx
Z 1
= φ(x)dx − φ(1) + 2φ(1)
0
Z 1
= φ(x)dx + φ(1) .
0
Suppose that φj is a sequence in C0∞ (0, 2) such that φj (1) = 1 and φj (x) → 0
for x 6= 1. Then
Z 2 Z 1
1 = φj (1) = v(x)φj (x)dx − φj (x)dx → 0 ,
0 0
which provides the contradiction.
Definition 2.67. For p ∈ [1, ∞], define W 1,p (Ω) = {u ∈ Lp (Ω) | weak derivative exists , Du ∈
Lp (Ω)}, where Du is the weak derivative of u.
Example 2.68. Let n = 1 and set Ω = (0, 1). Define the function f (x) =
du
sin(1/x). Then u ∈ L1 (0, 1) and dx = − cos(1/x)/x2 ∈ L1loc (0, 1), but u 6∈
W 1,p (Ω) for any p.
Definition 2.69. In the case p = 2, we set H 1 (Ω) = W 1,p (Ω).
Example 2.70. Let Ω = B(0, 1) ⊂ R2 and set u(x) = |x|−α . We want to

determine the values of α for which u ∈ H 1 (Ω).
Since |x|−α = 3j=1 (xj xj )−α/2 , then ∂xi |x|−α = −α|x|−α−2 xi is well-
P
defined away from x = 0.
Step 1. We show that u ∈ L1loc (Ω). To see this, note that −α dx

R
Ω |x| =
R 2π R 1 −α
0 0 r rdrdθ < ∞ whenever α < 2.
Step 2. Set the vector v(x) = −α|x|−α−2 x (so that each component is given
by vi (x) = −α|x|−α−2 xi ). We show that
Z Z
u(x)Dφ(x)dx = − v(x)φ(x)dx ∀φ ∈ C0∞ (B(0, 1)) .
B(0,1) B(0,1)
To see this, let Ωδ = B(0, 1) − B(0, δ), let n denote the unit normal to ∂Ωδ
(pointing toward the origin). Integration by parts yields
Z Z 2π Z
−α −α
|x| Dφ(x)dx = δ φ(x)n(x)δdθ + α |x|−α−2 x φ(x)dx .
Ωδ 0 Ωδ
R 2π
Since limδ→0 δ 1−α 0 φ(x)n(x)dθ = 0 if α < 1, we see that
Z Z
−α
lim |x| Dφ(x)dx = lim α |x|−α−2 x φ(x)dx
δ→0 Ωδ δ→0 Ωδ
R 2π R 1
Since 0 0 r−α−1 rdrdθ < ∞ if α < 1, the Dominated Convergence Theo-
rem shows that v is the weak derivative of u.
R 2π R 1
Step 3. v ∈ L2 (Ω), whenever 0 0 r−2α−2 rdrdθ < ∞ which holds if α < 0.
Remark 2.71. Note that if the weak derivative exists, it is unique. To see
this, suppose that both v1 and v2 are the weak derivative of u on Ω. Then
∞
R
Ω (v1 − v2 )φdx = 0 for all φ ∈ C0 (Ω), so that v1 = v2 a.e.
2.2.2 Definition of Sobolev Spaces
Definition 2.72. For integers k ≥ 0 and 1 ≤ p ≤ ∞,
W k,p (Ω) = {u ∈ L1loc (Ω) |Dα u exists and is in Lp (Ω) for |α| ≤ k}.
Definition 2.73. For u ∈ W k,p (Ω) define

 1
p
ukpLp (Ω) 
X
α
kukW k,p (Ω) =  kD for 1 ≤ p < ∞ ,
|α|≤k
and
X
kukW k,∞ (Ω) = kDα ukL∞ (Ω) .
|α|≤k
The function k · kW k,p (Ω) is clearly a norm since it is a finite sum of Lp

norms.
Definition 2.74. A sequence uj → u in W k,p (Ω) if limj→∞ kuj −ukW k,p (Ω) =
0.
Theorem 2.75. W k,p (Ω) is a Banach space.
Proof. Let uj denote a Cauchy sequence in W k,p (Ω). It follows that for all
|α| ≤ k, Dα uj is a Cauchy sequence in Lp (Ω). Since Lp (Ω) is a Banach
space (see Theorem 2.19), for each α there exists uα ∈ Lp (Ω) such that
Dα uj → uα in Lp (Ω) .
When α = (0, ..., 0) we set u := u(0,...,0) so that uj → u in Lp (Ω). We must

show that uα = Dα u.
For each φ ∈ C0∞ (Ω),
Z Z
α
uD φdx = lim uj Dα φdx
Ω j→∞ Ω
Z
|α|
= (−1) lim Dα uj φdx
j→∞ Ω
Z
= (−1)|α| uα φdx ;
Ω
thus, uα = Dα u and hence Dα uj → Dα u in Lp (Ω) for each |α| ≤ k, which

shows that uj → u in W k,p (Ω).
Definition 2.76. For integers k ≥ 0 and p = 2, we define
H k (Ω) = W k,2 (Ω) .
H k (Ω) is a Hilbert space with inner-product (u, v)H k (Ω) = α u, D α v)

P
|α|≤k (D L2 (Ω) .
2.2.3 A simple version of the Sobolev embedding theorem
For two Banach spaces B1 and B2 , we say that B1 is embedded in B2 if

kukB2 ≤ CkukB1 for some constant C and for u ∈ B1 . We wish to determine
which Sobolev spaces W k,p (Ω) can be embedded in the space of continuous
functions. To motivate the type of analysis that is to be employed, we study
a special case.
Theorem 2.77 (Sobolev embedding in 2-D). For kp ≥ 2,
max |u(x)| ≤ CkukW k,p (R2 ) ∀u ∈ C0∞ (Ω) . (2.7)

x∈R2
Proof. Given u ∈ C0∞ (Ω), we prove that for all x ∈ spt(u),
|u(x)| ≤ CkDα u(x)kLp (Ω) ∀|α| ≤ k .
By choosing a coordinate system centered about x, we can assume that

x = 0; thus, it suffices to prove that
|u(0)| ≤ CkDα u(x)kLp (Ω) ∀|α| ≤ k .
Let g ∈ C ∞ ([0, ∞)) with 0 ≤ g ≤ 1, such that g(x) = 1 for x ∈ [0, 21 ] and
g(x) = 0 for x ∈ [ 34 , ∞).
By the fundamental theorem of calculus,
Z 1 Z 1
u(0) = − ∂r [g(r)u(r, θ)]dr = − ∂r (r) ∂r [g(r)u(r, θ)]dr
0 0
Z 1
= r ∂r2 [g(r)u(r, θ)]dr
0
Z 1 Z 1
(−1)k k−1 k (−1)k
= r ∂r [g(r)u(r, θ)]dr = rk−2 ∂rk [g(r)u(r, θ)]rdr
(k − 1)! 0 (k − 1)! 0
Integrating both sides from 0 to 2π, we see that

2π 1
(−1)k
Z Z
u(0) = rk−2 ∂rk [g(r)u(r, θ)]rdrdθ .
2π(k − 1)! 0 0
The change of variables from Cartesian to polar coordinates is given by
x(r, θ) = r cos θ , y(r, θ) = r sin θ .

By the chain-rule,
∂r u(x(r, θ), y(r, θ)) = ∂x u cos θ + ∂y u sin θ ,

∂r2 u(x(r, θ), y(r, θ)) = ∂x2 u cos2 θ + 2∂xy
2
u cos θ sin θ + ∂y2 u sin2 θ
..
.
It follows that ∂rk = aα (θ)Dα , where aα consists of trigonometric

P
|α|≤k
polynomials of θ, so that
(−1)k
Z X
u(0) = rk−2 aα (θ)Dα [g(r)u(x)]dx
2π(k − 1)! B(0,1) |α|≤k
X
≤ Ckrk−2 kLq (B(0,1)) kDα (gu)kLp (B(0,1))
|α|≤k
Z 1 p−1
p
p(k−2)
≤C r p−1 rdr kukW k,p (R2 ) .
0
p(k−2)
Hence, we require p−1 + 1 > −1 or kp > 2.
2.2.4 Approximation of W k,p (Ω) by smooth functions
Recall that Ω = {x ∈ Ω | dist(x, ∂Ω) > }.
Theorem 2.78. For integers k ≥ 0 and 1 ≤ p < ∞, let
u = η ∗ u in Ω ,
where η is the standard mollifier defined in Definition 2.25. Then
(A) u ∈ C ∞ (Ω ) for each > 0, and
k,p
(B) u → u in Wloc (Ω) as → 0.
k,p
Definition 2.79. A sequence uj → u in Wloc (Ω) if uj → u in W k,p (Ω̃) for
each Ω̃ ⊂⊂ Ω.
Proof of Theorem 2.78. Theorem 2.28 proves part (A). Next, let v α denote
the the αth weak partial derivative of u. To prove part (B), we show that
Dα u = η ∗ v α in Ω . For x ∈ Ω ,
Z
α α
D u (x) = D η (x − y)u(y)dy
Z Ω
= Dxα η (x − y)u(y)dy
Ω Z
|α|
= (−1) Dyα η (x − y)u(y)dy
Z Ω
= η (x − y)v α (y)dy = (η ∗ v α )(x) .

Ω
By part (D) of Theorem 2.28, Dα u → v α in Lploc (Ω).
It is possible to refine the above interior approximation result all the

way to the boundary of Ω. We record the following theorem without proof.
Theorem 2.80. Suppose that Ω ⊂ Rn is a smooth, open, bounded subset,

and that u ∈ W k,p (Ω) for some 1 ≤ p < ∞ and integers k ≥ 0. Then there
exists a sequence uj ∈ C ∞ (Ω) such that
uj → u in W k,p (Ω) .
It follows that the inequality (2.7) holds for all u ∈ W k,p (R2 ).
2.2.5 Hölder Spaces
Recall that for Ω ⊂ Rn open and smooth, the class of Lipschitz functions
u : Ω → R satisfies the estimate
|u(x) − u(y)| ≤ C|x − y| ∀x, y ∈ Ω
for some constant C.
Definition 2.81 (Classical derivative). A function u : Ω → R is differen-

tiable at x ∈ Ω if there exists f : Ω → L(Rn ; Rn ) such that
|u(x) − u(y) − f (x) · (x − y)|

→ 0.
|x − y|
We call f (x) the classical derivative (or gradient) of u(x), and denote it by
Du(x).
Definition 2.82. If u : Ω → R is bounded and continuous, then
kukC 0 (Ω) = max |u(x)| .

x∈Ω
If in addition u has a continuous and bounded derivative, then
kukC 1 (Ω) = kukC 0 (Ω) + kDukC 0 (Ω) .
The Hölder spaces interpolate between C 0 (Ω) and C 1 (Ω).
Definition 2.83. For 0 < γ ≤ 1, the space C 0,γ (Ω) consists of those func-
tions for which
kukC 0,γ (Ω) := kukC 0 (Ω) + [u]C 0,γ (Ω) < ∞ ,
where the γth Hölder semi-norm [u]C 0,γ (Ω) is defined as
|u(x) − u(y)|

[u]C 0,γ (Ω) = max .
x,y∈Ω |x − y|γ
x6=y
The space C 0,γ (Ω) is a Banach space.
2.2.6 Morrey’s inequality
We can now offer a refinement and extension of the simple version of the
Sobolev Embedding Theorem 2.77.
Theorem 2.84 (Morrey’s inequality). For n < p ≤ ∞, let B(x, r) ⊂ Rn

and let y ∈ B(x, r). Then
1− n
|u(x) − u(y)| ≤ Cr p kDukLp (B(x,2r)) ∀u ∈ C 1 (Rn ) .
In fact, Morrey’s inequality holds for all u ∈ W 1,p (B(x, 2r)) (see Problem
2.25 in the Exercises).
Notation 2.85 (Averaging). Let B(0, 1) ⊂ Rn . The volume of B(0, 1) is

n
π2
given by αn = Γ( n +1) and the surface area is |Sn−1 | = nαn . We define
2
1
Z Z
− f (y)dy = f (y)dy
B(x,r) αn rn B(x,r)
1
Z Z
− f (y)dS = f (y)dS .
∂B(x,r) nαn rn−1 ∂B(x,r)
Lemma 2.86. For B(x, r) ⊂ Rn , y ∈ B(x, r) and u ∈ C 1 (B(x, r)),
|Du(y)|
Z Z
− |u(y) − u(x)|dy ≤ C n−1
dy .
B(x,r) B(x,r) |x − y|
Proof. For some 0 < s < r, let y = x + sω where ω ∈ Sn−1 = ∂B(0, 1). By
the fundamental theorem of calculus, for 0 < s < r,
Z s
d
u(x + sω) − u(x) = u(x + tω)dt
dt
Z0 s
= Du(x + tω) ωdt .
0
Since |ω| = 1, it follows that

Z s
|u(x + sω) − u(x)| ≤ |Du(x + tω)|dt .
0
Thus, integrating over Sn−1 yields

Z Z sZ
|u(x + sω) − u(x)|dω ≤ |Du(x + tω)|dωdt
Sn−1 0 Sn−1
Z sZ
tn−1
≤ |Du(x + tω)| n−1 dωdt
n−1 t
Z0 S
|Du(y)|
≤ n−1
dy ,
B(x,r) |x − y|
where we have set y = x + tω.

Multipling the above inequality by sn−1 and integrating s from 0 to r
shows that
Z rZ
rn |Du(y)|
Z
n−1
|u(x + sω) − u(x)|dωs ds ≤ dy
0 Sn−1 n B(x,r) |x − y|n−1
|Du(y)|
Z
n
≤ Cαn r dy ,
B(x,r) |x − y|n−1
which proves the lemma.
Proof of Theorem 2.84. Assume first that u ∈ C 1 (B(x, 2r)). Let D =

B(x, r) ∩ B(y, r) and set r = |x − y|. Then
Z
|u(x) − u(y)| = − |u(x) − u(y)|dz
ZD Z
≤ − |u(x) − u(z)|dz + − |u(y) − u(z)|dz .
D D
Since D equals the intersection of two balls of radius r, it is clear that can
choose a constant C, depending only on the dimension n, such that
|D| |D|
= =C.
|B(x, r)| |B(y, r)|
It follows that
Z Z
|u(x) − u(y)| ≤ − |u(x) − u(z)|dz + − |u(y) − u(z)|dz
D D
Z
C
Z
≤ |u(x) − u(z)|dz + |u(y) − u(z)|dz
|B(x, r)| D D
Z Z
≤ C− |u(x) − u(z)|dz + C− |u(y) − u(z)|dz .
B(x,r) B(y,r)
Thus, by Lemma 2.86,

Z Z Z
1−n
− |u(x)−u(z)|dz ≤ C |x−z| |Du(z)|dz ≤ C |x−z|1−n |Du(z)|dz
B(x,r) B(x,r) B(x,2r)
and
Z Z Z
1−n
− |u(x)−u(z)|dz ≤ C |x−z| |Du(z)|dz ≤ C |x−z|1−n |Du(z)|dz
B(y,r) B(y,r) B(x,2r)
so that
Z
|u(x) − u(y)| ≤ C |x − z|1−n |Du(z)|dz (2.8)
B(x,2r)
and by Hölder’s inequality,

! p−1 !1
Z p Z p
p(1−n)
n−1 p
|u(x) − u(y)| ≤ C s p−1 s dsdω |Du(z)| dz
B(0,2r) B(x,2r)
Morrey’s inequality implies the following embedding theorem.
Theorem 2.87 (Sobolev embedding theorem for k = 1). There exists a

constant C = C(p, n) such that
kuk 0,1− n ≤ CkukW 1,p (Rn ) ∀u ∈ W 1,p (Rn ) .

C p (Rn )
Proof. First assume that u ∈ C01 (Rn ). Given Morrey’s inequality, it suffices
to show that max |u| ≤ CkukW 1,p (Rn ) . Using Lemma 2.86, for all x ∈ Rn ,
Z Z
|u(x)| ≤ − |u(x) − u(y)|dy + − |u(y)|dy
B(x,1) B(x,1)
|Du(y)|
Z
≤C dy + CkukLp (Rn )
B(x,1) |x − y|n−1
≤ CkukW 1,p (Rn ) ,
the last inequality following whenever p > n.

Thus,
kuk 0,1− n ≤ CkukW 1,p (Rn ) ∀u ∈ C 1 (Rn ) . (2.9)
C p (Rn )
By the density of C0∞ (Rn ) in W 1,p (Rn ), there is a sequence uj ∈ C0∞ (Rn )
such that
uj → u ∈ W 1,p (Rn ) .
By (2.9), for j, k ∈ N,
0,1− n
kuj − uk kC p (Rn ) ≤ Ckuj − uk kW 1,p (Rn ) .
0,1− n 0,1− n
Since C p (Rn ) is a Banach space, there exists a U ∈ C p (Rn ) such
that
0,1− n
uj → U in C p (Rn ) .
It follows that U = u a.e. in Ω. By the continuity of norms with respect to

strong convergence, we see that
kU k 0,1− n ≤ CkukW 1,p (Rn )

C p (Rn )
which completes the proof.
In proving the above embedding theorem, we established that that for

p > n, we have the inequality
kukL∞ (Rn ) ≤ CkukW 1,p (Rn ) . (2.10)
We will see later that (2.10), via a scaling argument, leads to the following
important interpolation inequality: for p > n,
n p−n
kukL∞ (Rn ) ≤ C(n, p)kDukLp p (Rn ) kukLpp(Rn ) .
Another important consequence of Morrey’s inequality is the relation-

ship between the weak and classical derivative of a function. We begin by
recalling the definition of classical differentiability. A function u : Rn → Rm
is differentiable at a point x if there exists a linear operator L : Rn → Rm
such that for each > 0, there exists δ > 0 with |y − x| < δ implying that
ku(y) − u(x) − L(y − x)k ≤ ky − xk .
When such an L exists, we write Du(x) = L and call it the classical deriva-
tive.
As a consequence of Morrey’s inequality, we extract information about
the classical differentiability properties of weak derivatives.
Theorem 2.88 (Differentiability a.e.). If Ω ⊂ Rn , n < p ≤ ∞ and u ∈

1,p
Wloc (Ω), then u is differentiable a.e. in Ω, and its gradient equals its weak
gradient almost everywhere.
Proof. We first restrict n < p < ∞. By a version Lebesgue’s differentiation

theorem, for almost every x ∈ Ω,
Z
lim − |Du(x) − Du(z)|p dz = 0 , (2.11)
r→0 B(x,r)
where Du denotes the weak derivative of u. Thus, for r > 0 sufficiently

small, we see that
Z
− |Du(x) − Du(z)|p dz < .
B(x,r)
Fix a point x ∈ Ω for which (2.11) holds, and define the function
wx (y) = u(y) − u(x) − Du(x) · (y − x) .
Notice that wx (x) = 0 and that
Dy wx (y) = Du(y) − Du(x) .
Set r = |x − y|. Since |u(y) − u(x) − Du(x) · (y − x)| = |wx (y) − wx (x)|, an
application of the inequality (2.8) that we obtained in the proof of Morrey’s
inequality then yields the estimate
|Dz wx (z)|
Z
|u(y) − u(x) − Du(x) · (y − x)| ≤ C dz
B(x,2r) |x − z|n−1
|Du(z) − Du(x)|
Z
=C dz
B(x,2r) |x − z|n−1
!1
Z p
1− n
≤ Cr p |Du(z) − Du(x)|p dz
B(x,2r)
!1
Z p
≤ Cr − |Du(z) − Du(x)|p dz
B(x,2r)
≤ C|x − y| ,
from which it follows that Du(x) is the classical derivative of u at the point
x.
1,∞ 1,p
The case that p = ∞ follows from the inclusion Wloc (Ω) ⊂ Wloc (Ω) for
all 1 ≤ p < ∞.
2.2.7 The Gagliardo-Nirenberg-Sobolev inequality
In the previous section, we considered the embedding for the case that p > n.
Theorem 2.89 (Gagliardo-Nirenberg inequality). For 1 ≤ p < n, set p∗ =

np
n−p . Then
kukLp∗ (Rn ) ≤ Cp,n kDukLp (Rn ) ∀u ∈ W 1,p (Rn ) .
Proof for the case n = 2. Suppose first that p = 1 in which case p∗ = 2, and
we must prove that
kukL2 (R2 ) ≤ CkDukL1 (R2 ) ∀u ∈ C01 (R2 ) . (2.12)
Since u has compact support, by the fundamental theorem of calculus,

Z x1 Z x2
u(x1 , x2 ) = ∂1 u(y1 , x2 )dy1 = ∂2 u(x1 , y2 )dy2
−∞ −∞
so that
Z ∞ Z ∞
|u(x1 , x2 )| ≤ |∂1 u(y1 , x2 )|dy1 ≤ |Du(y1 , x2 )|dy1
−∞ −∞
and
Z ∞ Z ∞
|u(x1 , x2 )| ≤ |∂2 u(x1 , y2 )|dy2 ≤ |Du(x1 , y2 )|dy2 .
−∞ −∞
Hence, it follows that

Z ∞ Z ∞
2
|u(x1 , x2 )| ≤ |Du(y1 , x2 )|dy1 |Du(x1 , y2 )|dy2
−∞ −∞
Integrating over R2 , we find that

Z ∞Z ∞
|u(x1 , x2 )|2 dx1 dx2
−∞ −∞
Z ∞ Z ∞ Z ∞ Z ∞
≤ |Du(y1 , x2 )|dy1 |Du(x1 , y2 )|dy2 dx1 dx2
−∞ −∞ −∞ −∞
Z ∞ Z ∞ 2
≤ |Du(x1 , x2 )|dx1 dx2
−∞ −∞
which is (2.12).
Next, if 1 ≤ p < 2, substitute |u|γ for u in (2.12) to find that
Z 1 Z
2
2γ
|u| dx ≤ Cγ |u|γ−1 |Du|dx
R2 R2
Z p−1
p(γ−1) p
≤ CγkDukLp (R2 ) |u| p−1 dx
R2
p(γ−1) p
Choose γ so that 2γ = p−1 ; hence, γ = 2−p , and
Z 2−p
2p 2p
|u| 2−p dx ≤ CγkDukLp (R2 ) ,
R2
so that
kuk 2p ≤ Cp,n kDukLp (Rn ) (2.13)
L 2−p (Rn )
for all u ∈ C01 (R2 ).

Since C0∞ (R2 ) is dense in W 1,p (R2 ), there exists a sequence uj ∈ C0∞ (R2 )
such that
uj → u in W 1,p (R2 ) .
Hence, by (2.13), for all j, k ∈ N,
kuj − uk k 2p ≤ Cp,n kDuj − Duk kLp (Rn )

L 2−p (Rn )
2p
so there exists U ∈ L 2−p (Rn ) such that
2p
uj → U in L 2−p (Rn ) .
Hence U = u a.e. in R2 , and by continuity of the norms, (2.13) holds for all
u ∈ W 1,p (R2 ).
Proof for the general case of dimension n. Following the proof for n = 2,
we see that
1
n
Z ∞
n−1
|u(x)| n−1 ≤ Πni=1 |Du(x1 , ..., yi , ..., xn )|dyi
−∞
so that
1
Z ∞ n
Z ∞ Z ∞
n−1
|u(x)| n−1 dx1 ≤ Πni=1 |Du(x1 , ..., yi , ..., xn )|dyi dx1
−∞ −∞ −∞
1 1
Z ∞
n−1
Z ∞ Z ∞
n−1
|Du|dy1 Πni=2 |Du|dyi dx1
−∞ −∞ −∞
1 1
Z ∞
n−1
Z ∞ Z ∞
n−1
|Du|dy1 Πni=2 |Du|dx1 dyi ,
−∞ −∞ −∞
where the last inequality follows from Hölder’s inequality.

Integrating the last inequality with respect to x2 , we find that
1
Z ∞ Z ∞ n
Z ∞ Z ∞
n−1
Z ∞ 1
|u(x)| n−1 dx1 dx2 < |Du|dx1 dy2 Πni=1 Iin−1 dx2 ,
−∞ −∞ −∞ −∞ −∞ i6=2
where
Z ∞ Z ∞ Z ∞
I1 = |Du|dy1 , Ii = |Du|dx1 dyi for i = 3, ..., n .
−∞ −∞ −∞
Applying Hölder’s inequality, we find that

Z ∞Z ∞
n
|u(x)| n−1 dx1 dx2
−∞ −∞
1 1
Z ∞ Z ∞
n−1
Z ∞ Z ∞
n−1
≤ |Du|dx1 dy2 |Du|dy1 dx2
−∞ −∞ −∞ −∞
1
Z ∞ Z ∞ Z ∞
n−1
Πni=3 |Du|dx1 dx2 dyi .
−∞ −∞ −∞
Next, continue to integrate with respect to x3 , ..., xn to find that

1
Z
n
Z ∞ Z ∞
n−1
|u| n−1 dx ≤ Πni=1 ··· |Du|dx1 ...dyi ...dxn
Rn −∞ −∞
Z n
n−1
= |Du|dx .
Rn
This proves the case that p = 1. The case that 1 < p < n follows identically
as in the proof of n = 2.
It is common to employ the Gagliardo-Nirenberg inequality for the case

that p = 2; as stated, the inequality is not well-defined in dimension two,
but in fact, we have the following theorem.
Theorem 2.90. Suppose that u ∈ H 1 (R2 ). Then for all 1 ≤ q < ∞,
√
kukLq (R2 ) ≤ C qkukH 1 (R2 ) .
Proof. Let x and y be points in R2 , and write r = |x − y|. Let θ ∈ S1 .

Introduce spherical coordinates (r, θ) with origin at x, and let g be the
same cut-off function that was used in the proof of Theorem 2.77. Define
U := g(r)u(r, θ). Then
1 1
∂U ∂U
Z Z
u(x) = − (r, θ)dr − |x − y|−1 (r, θ)rdr
0 ∂r 0 ∂r
and
Z 1
|u(x)| ≤ |x − y|−1 |DU (r, θ)|rdr .
0
Integrating over S1 , we obtain:
1
Z
|u(x)| ≤ 1B(x,1) |x − y|−1 |DU (y)|dy := K ∗ |DU | ,
2π R2
1 −1
where the integral kernel K(x) = 2π 1B(0,1) |x| .
Using Young’s inequality from Theorem 2.54, we obtain the estimate
1 1 1
kK ∗ f kLq (R2 ) ≤ kKkLk (R2 ) kf kL2 (R2 ) for = − + 1. (2.14)
k q 2
Using the inequality (2.14) with f = |DU |, we see that

"Z #1
k
kukLq (R2 ) ≤ CkDU kL2 (R2 ) |y|−k dy

B(0,1)
Z 1 k1
1−k
≤ CkDU kL2 (R2 ) r dr
0
1
q+2 k
= CkukH 1 (R2 ) .
4
1
When q → ∞, k → 21 , so
1
kukLq (R2 ) ≤ Cq 2 kukH 1 (R2 ) .
Evidently, it is not possible to obtain the estimate kukL∞ (Rn ) ≤ CkukW 1,n (Rn )
with a constant C < ∞. The following provides an example of a function in
this borderline situation.
Example 2.91. Let Ω ⊂ R2 denote the open unit ball in R2 . The unbounded

1
function u = log log 1 + |x| belongs to H 1 (B(0, 1)).
First, note that
2π 1h
1 i2
Z Z Z
|u(x)|2 dx = log log 1 + rdrdθ .
Ω 0 0 r
The only potential singularity of the integrand occurs at r = 0, but according

to L’Hospital’s rule,
h 1 i2
lim r log log 1 + = 0, (2.15)
r→0 r
so the integrand is continuous and hence u ∈ L2 (Ω).

In order to compute the partial derivatives of u, note that
df
∂ xj d f (x) dz
|x| = , and |f (z)| = ,
∂xj |x| dz |f (z)|
where f : R → R is differentiable. It follows that for x away from the origin,
−x
Du(x) = 1 , (x 6= 0) .
log(1 + |x| )(|x| + 1)|x|2
Let φ ∈ C0∞ (Ω) and fix > 0. Then

∂φ ∂u
Z Z Z
u(x) (x)dx = − (x)φ(x)dx + uφNi dS ,
Ω−B (0) ∂xi Ω−B(0,) ∂xi ∂B(0,)
where N = (N1 , ..., Nn ) denotes the inward-pointing unit normal on the

curve ∂B(0, ), so that N dS = (cos θ, sin θ)dθ. It follows that
Z Z
u(x)Dφ(x)dx = − Du(x)φ(x)dx
Ω−B (0) Ω−B (0)
2π
1
Z
− (cos θ, sin θ) log log 1 + φ(, θ)dθ . (2.16)
0
We claim that Du ∈ L2 (Ω) (and hence also in L1 (Ω)), for

2π 1
1
Z Z Z
2
|Du(x)| dx = h i2 drdθ
Ω 0 0 r(r + 1)2 log 1 + 1r
1/2 Z 1
1 1
Z
≤ π 2
dr + π i2 dr
r(log r)
h
0 1/2 r(r + 1)2 log 1 + 1
r
where we use the inequality log(1 + 1r ) ≥ log 1r = − log r ≥ 0 for 0 ≤ r ≤ 1.

The second integral on the right-hand side is clearly bounded, while
1/2 − log 2 − log 2
1 1 t 1
Z Z Z
dr = e dt = dx < ∞ ,
0 r(log r)2 −∞ t et
2
−∞ x2
so that Du ∈ L2 (Ω). Letting → 0 in (2.16) and using (2.15) for the

boundary integral, by the Dominated Convergence Theorem, we conclude
that Z Z
u(x)Dφ(x)dx = − Du(x)φ(x)dx ∀φ ∈ C0∞ (Ω) .
Ω Ω
2.2.8 Local coordinates near ∂Ω
Let Ω ⊂ Rn denote an open, bounded subset with C 1 boundary, and let

{Ul }K
l=1 denote an open covering of ∂Ω, such that for each l ∈ {1, 2, ..., K},
with
Vl = B(0, rl ), denoting the open ball of radius rl centered at the origin and,
Vl+ = Vl ∩ {xn > 0} ,
Vl− = Vl ∩ {xn < 0} ,
there exist C 1 -class charts θl which satisfy
θl : Vl → Ul is a C 1 diffeomorphism , (2.17)
θl (Vl+ ) = Ul ∩ Ω ,
θl (Vl ∩ {xn = 0}) = Ul ∩ ∂Ω .
2.2.9 Sobolev extensions and traces.
Let Ω ⊂ Rn denote an open, bounded domain with C 1 boundary.
Theorem 2.92. Suppose that Ω̃ ⊂ Rn is a bounded and open domain such

that Ω ⊂⊂ Ω̃. Then for 1 ≤ p ≤ ∞, there exists a bounded linear operator
E : W 1,p (Ω) → W 1,p (Rn )
such that for all u ∈ W 1,p (Ω),
1. Eu = u a.e. in Ω;
2. spt(Eu) ⊂ Ω̃;
3. kEukW 1,p (Rn ) ≤ CkukW 1,p (Ω) for a constant C = C(p, Ω, Ω̃).
Theorem 2.93. For 1 ≤ p < ∞, there exists a bounded linear operator
T : W 1,p (Ω) → Lp (Ω)
such that for all u ∈ W 1,p (Ω)
1. T u = u|∂Ω for all u ∈ W 1,p (Ω) ∪ C 0 (Ω);
2. kT ukLp (∂Ω) ≤ CkukW 1,p (Ω) for a constant C = C(p, Ω).
Proof. Suppose that u ∈ C 1 (Ω), z ∈ ∂Ω, and that ∂Ω is locally flat near
z. In particular, for r > 0 sufficiently small, B(z, r) ∪ ∂Ω ⊂ {xn = 0}. Let
0 ≤ ξ ∈ C0∞ (B(z, r)) such that ξ = 1 on B(z, r/2). Set Γ = ∂Ω ∪ B(z, r/2),
B + (z, r) = B(z, r) ∪ Ω, and let dxh = dx1 · · · dxn−1 . Then

Z Z
p
|u| dxh ≤ ξ|u|p dxh
Γ {xn =0}
∂
Z
=− (ξ|u|p )dx
+
B (z,r) ∂x n
∂ξ ∂u
Z Z
p
≤− |u| dx − p ξ|u|p−2 u dx
B + (z,r) ∂xn B + (z,2δ) ∂xn

∂u
Z
p p−1
≤C |u| dx + Ck|u| k p−1 p
B + (z,r) L (B + (z,r)) ∂xn p +
L (B (z,r))
Z
≤C (|u|p + |Du|p )dx . (2.18)
B + (z,r)
On the other hand, if the boundary is not locally flat near z ∈ ∂Ω,
then we use a C 1 diffeomorphism to locally straighten the boundary. More
specifically, suppose that z ∈ ∂Ω ∪ Ul for some l ∈ {1, ..., K} and consider
the C 1 chart θl defined in (2.17). Define the function U = u ◦ θl ; then
U : Vl+ → R. Setting Γ = Vl ∪ {xn = 0k, we see from the inequality (2.18),
that
Z Z
p
|U | dxh ≤ Cl (|U |p + |DU |p )dx .
Γ Vl+
Using the fact that Dθl is bounded and continuous on Vl+ , the change of
variables formula shows that
Z Z
|u|p dS ≤ Cl (|u|p + |Du|p )dx .
Ul ∪∂Ω Ul+
Summing over all l ∈ {1, ..., K} shows that

Z Z
|u| dS ≤ C (|u|p + |Du|p )dx .
p
(2.19)
∂Ω Ω
The inequality (2.19) holds for all u ∈ C 1 (Ω). According to Theorem 2.80,
for u ∈ W 1,p (Ω) there exists a sequence uj ∈ C ∞ (Ω) such that uj → u in
W 1,p (Ω). By inequality (2.19),
kT uk − T uj kLp (∂Ω) ≤ Ckuk − uj kW 1,p (Ω) ,
so that T uj is Cauchy in Lp (∂Ω), and hence a limit exists in Lp (∂Ω) We

define the trace operator T as this limit:
lim kT u − T uj kLp (∂Ω) = 0 .

j→0
Since the sequence uj converges uniformly to u if u ∈ C 0 (Ω), we see that

T u = u|∂Ω for all u ∈ W 1,p (Ω) ∪ C 0 (Ω).
Sketch of the proof of Theorem 2.92. Just as in the proof of the trace the-
orem, first suppose that u ∈ C 1 (ω) and that near z ∈ ∂Ω, ∂Ω is lo-
cally flat, so that for some r > 0, ∂Ω ∪ B(z, r) ⊂ {xn = 0}. Letting
B + = B(z, r) ∪ {xn ≥ 0} and B − = B(z, r) ∪ {xn ≤ 0} , we define the
extension of u by
if x ∈ B +

u(x)
ū(x) =
−3u(x1 , ..., xn−1 , −xn ) + 4u(x1 , ..., xn−1 , − x2n ) if x ∈ B − .
Define u+ = ū|B + and u− = ū|B − .
It is clear that u+ = u− on {xn = 0}, and by the chain-rule, it follows
that
∂u− ∂u− ∂u− xn
(x) = 3 (x1 , ..., −xn ) − 2 (x1 , ..., − ) ,
∂xn ∂xn ∂xn 2
∂u+ ∂u−
so that ∂xn = ∂xn on {xn = 0}. This shows that ū ∈ C 1 (B(z, r). using the
charts θl to locally straighten the boundary, and the density of the C ∞ (Ω)
in W 1,p (Ω), the theorem is proved.
Later, we will provide a proof for higher-order Sobolev extensions of

H k -type functions.
2.2.10 The subspace W01,p (Ω)
Definition 2.94. We let W01,p (Ω) denote the closure of C0∞ (Ω) in W 1,p (Ω).
Theorem 2.95. Suppose that Ω ⊂ Rn is bounded with C 1 boundary, and

that u ∈ W 1,p (Ω). Then
u ∈ W01,p (Ω) iff T u = 0 on ∂Ω .
We can now state the Sobolev embedding theorems for bounded domains
Ω.
Theorem 2.96 (Gagliardo-Nirenberg inequality for W 1,p (Ω)). Suppose that

Ω ⊂ Rn is open and bounded with C 1 boundary, 1 ≤ p < n, and u ∈ W 1,p (Ω).
Then
kuk np ≤ CkukW 1,p (Ω) for a constant C = C(p, n, Ω) .

L n−p (Ω)
Proof. Choose Ω̃ ⊂ Rn bounded such that Ω ⊂⊂ Ω̃, and let Eu denote the
Sobolev extension of u to Rn such that Eu = u a.e., spt(Eu) ⊂ Ω̃, and
kEukW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
Then by the Gagliardo-Nirenberg inequality,
kuk np ≤ kEuk np ≤ CkD(Eu)kLp (Rn )

L n−p (Ω) L n−p (Rn )
≤ CkEukW 1,p (Rn ) ≤ CkukW 1,p (Ω) .
By following the proof of Theorem 2.89, we have the following general-

ization for integers k ≥ 1:
Theorem 2.97 (Gagliardo-Nirenberg inequality for W k,p (Ω)). Suppose that

Ω ⊂ Rn is open and bounded with C 1 boundary, 1 ≤ kp < n, and u ∈
W k,p (Ω). Then
kuk np ≤ CkukW k,p (Ω) for a constant C = C(k, p, n, Ω) .

L n−pk (Ω)
In fact, the theorem is true for real numbers s > 0 replacing integers k ≥
1, and follows from linear interpolation and the theory of fractional-order
Sobolev spaces defined later in Section 2.5.2. In the important case that
p = 2, we are then able to answer the question of which H s spaces embed
in Lq spaces. For example, when n = 2 and s = 12 , we see that kukL4 (Ω) ≤
Ckuk 1 , and when n = 3 and s = 21 , kuk 12 ≤ Ckuk 1 .
H 2 (Ω) L 5 (Ω) H 2 (Ω)
Theorem 2.98 (Gagliardo-Nirenberg inequality for W01,p (Ω)). Suppose that

Ω ⊂ Rn is open and bounded with C 1 boundary, 1 ≤ p < n, and u ∈ W01,p (Ω).
np
Then for all 1 ≤ q ≤ n−p ,
kukLq (Ω) ≤ CkDukLp (Ω) for a constant C = C(p, n, Ω) . (2.20)
Proof. By definition there exists a sequence uj ∈ C0∞ (Ω) such that uj → u

in W 1,p (Ω). Extend each uj by 0 on Ωc . Applying Theorem 2.89 to this
extension, and using the continuity of the norms, we obtain kuk pn ≤
L n−p (Ω)
CkDukLp (Ω) . Since Ω is bounded, the assertion follows by Hölder’s inequal-
ity.
Theorem 2.99. Suppose that Ω ⊂ R2 is open and bounded with C 1 bound-

ary, and u ∈ H01 (Ω). Then for all 1 ≤ q < ∞,
√
kukLq (Ω) ≤ C qkDukL2 (Ω) for a constant C = C(Ω) . (2.21)
Proof. The proof follows that of Theorem 2.90. Instead of introducing the
cut-off function g, we employ a partition of unity subordinate to the finite
covering of the bounded domain Ω, in which case it suffices that assume that
spt(u) ⊂ spt(U ) with U also defined in the proof Theorem 2.90.
Remark 2.100. Inequalities (2.20) and (2.21) are commonly referred to

as Poincaré inequalities, and will be discussed in further detail in Theorem
5.22. They are invaluable in the study of the Dirichlet problem for Poisson’s
equation, since the right-hand side provides an H 1 (Ω)-equivalent norm for
all u ∈ H01 (Ω). In particular, there exists constants C1 , C2 such that
C1 kDukL2 (Ω) ≤ kukH 1 (Ω) ≤ C2 kDukL2 (Ω) .
2.2.11 Weak solutions to Dirichlet’s problem
Suppose that Ω ⊂ Rn is an open, bounded domain with C 1 boundary. A

classical problem in the linear theory of partial differential equations consists
of finding solutions to the Dirichlet problem:
−∆u = f in Ω , (2.22a)
u = 0 on ∂Ω , (2.22b)
Pn ∂2
where ∆ = i=1 ∂x2i denotes the Laplace operator or Laplacian. As written,
(2.22) is the so-called strong form of the Dirichlet problem, as it requires
that u to possess certain weak second-order partial derivatives. A major
turning-point in the modern theory of linear partial differential equations
was the realization that weak solutions of (2.22) could be defined, which only
require weak first-order derivatives of u to exist. (We will see more of this
idea later when we discuss the theory of distributions.)
Definition 2.101. The dual space of H01 (Ω) is denoted by H −1 (Ω). For
f ∈ H −1 (Ω),
kf kH −1 (Ω) = sup hf, ψi ,
kψkH 1 (Ω) =1
0
where hf, ψi denotes the duality pairing between H −1 (Ω) and H01 (Ω).
Definition 2.102. A function u ∈ H01 (Ω) is a weak solution of (2.22) if

Z
Du · Dv dx = hf, vi ∀v ∈ H01 (Ω) .
Ω
Remark 2.103. Note that f can be taken in H −1 (Ω). According to the

Sobolev embedding theorem, this implies that when n = 1, the forcing func-
tion f can be taken to be the Dirac Delta distribution.
Remark 2.104. The motivation for Definition 2.102 is as follows. Since

C0∞ (Ω) is dense in H01 (Ω), multiply equation (2.22a) by φ ∈ C0∞ (Ω), inte-
R
grate over Ω, and employ the integration-by-parts formula to obtain Ω Du ·
R
Dφ dx = Ω f φ dx; the boundary terms vanish because φ is compactly sup-
ported.
Theorem 2.105 (Existence and uniqueness of weak solutions). For any

f ∈ H −1 (Ω), there exists a unique weak solution to (2.22).
Proof. Using the Poincaré inequality, kDukL2 (Ω) is an H 1 -equivalent norm

for all u ∈ H01 (Ω), and (Du, Dv)L2 (Ω) defines the inner-product on H01 (Ω).
As such, according to the definition of weak solutions to (2.22), we are
seeking u ∈ H01 (Ω) such that
(u, v)H01 (Ω) = hf, vi ∀v ∈ H01 (Ω) . (2.23)
The existence of a unique u ∈ H01 (Ω) satisfying (2.23) is provided by the

Riesz representation theorem for Hilbert spaces.
Remark 2.106. Note that the Riesz representation theorem shows that there
exists a distribution, denoted by −∆u ∈ H −1 (Ω) such that
h−∆u, vi = hf, vi ∀v ∈ H01 (Ω) .
The operator −∆ : H01 (Ω) → H −1 (Ω) is thus an isomorphism.
A fundamental question in the theory of linear partial differential equa-

tions is commonly referred to as elliptic regularity, and can be explained as
follows: in order to develop an existence and uniqueness theorem for the
Dirichlet problem, we have significantly generalized the notion of solution

to the class of weak solutions, which permitted very weak forcing functions
in H −1 (Ω). Now suppose that the forcing function is smooth; is the weak
solution smooth as well? Furthermore, does the weak solution agree with
the classical solution? The answer is yes, and we will develop this regular-
ity theory in Chapter 2.7, where it will be shown that for integers k ≥ 2,
−∆ : H k (Ω) ∩ H01 (Ω) → H k−2 (Ω) is also an isomorphism. An important
consequence of this result is that (−∆)−1 : H k−2 (Ω) → H k (Ω) ∩ H01 (Ω) is
a compact linear operator, and as such has a countable set of eigenvalues, a
fact that is eminently useful in the construction of solutions for heat- and
wave-type equations.
For this reason, as well as the consideration of weak limits of nonlinear
combinations of sequences, we must develop a compactness theorem, which
generalizes the well-known Arzela-Ascoli theorem to Sobolev spaces.
2.2.12 Strong compactness
In Section 2.1.12, we defined the notion of weak converence and weak com-
pactness for Lp -spaces. Recall that for 1 ≤ p < ∞, a sequence uj ∈ Lp (Ω)
converges weakly to u ∈ Lp (Ω), denoted uj * u in Lp (Ω), if Ω uj vdx →
R
R q p
Ω uvdx for all v ∈ L (Ω), with q = p−1 . We can extend this definition to
Sobolev spaces.
Definition 2.107. For 1 ≤ p < ∞, uj * u in W 1,p (Ω) provided that

uj * u in Lp (Ω) and Duj * Du in Lp (Ω).
Alaoglu’s Lemma (Theorem 2.37) then implies the following theorem.
Theorem 2.108 (Weak compactness in W 1,p (Ω)). For Ω ⊂ Rn , suppose

that
sup kuj kW 1,p (Ω) ≤ M < ∞ for a constant M 6= M (j) .
Then there exists a subsequence ujk * u in W 1,p (Ω).
It turns out that weak compactness often does not suffice for limit pro-
cesses involving nonlinearities, and that the Gagliardo-Nirenberg inequality
can be used to obtain the following strong compactness theorem.
Theorem 2.109 (Rellich’s theorem on a bounded domain Ω). Suppose that

Ω ⊂ Rn is an open, bounded domain with C 1 boundary, and that 1 ≤ p < n.
np
Then W 1,p (Ω) is compactly embedded in Lq (Ω) for all 1 ≤ q < n−p , i.e. if
sup kuj kW 1,p (Ω) ≤ M < ∞ for a constant M 6= M (j) ,
then there exists a subsequence ujk → u in Lq (Ω). In the case that n = 2

and p = 2, H 1 (Ω) is compactly embedded in Lq (Ω) for 1 ≤ q < ∞.
In order to prove Rellich’s theorem, we need two lemmas.
Lemma 2.110 (Arzela-Ascoli Theorem). Suppose that uj ∈ C 0 (Ω), kuj kC 0 (Ω) ≤

M < ∞, and uj is equicontinuous. Then there exists a subsequence ujk → u
uniformly on Ω.
Lemma 2.111. Let 1 ≤ r ≤ s ≤ t ≤ ∞, and suppose that u ∈ Lr (Ω)∩Lt (Ω).

1 a 1−a
Then for s = r + t
kukLs (Ω) ≤ kukaLr (Ω) kuk1−a

Lt (Ω) .
Proof. By Hölder’s inequality,

Z Z
s
|u| dx = |u|as |u|(1−a)s dx
Ω Ω
Z as Z (1−a)s
r r t t
as as (1−a)s (1−a)s (1−a)s
≤ |u| dx |u| dx = kukas
Lr (Ω) kukLt (Ω) .
Ω Ω
Proof of Rellich’s theorem. Let Ω̃ ⊂ Rn denote an open, bounded domain

such that Ω ⊂⊂ Ω̃. By the Sobolev extension theorem, the sequence uj
satisfies spt(uj ) ⊂ Ω̃, and
sup kEuj kW 1,p (Rn ) ≤ CM .
Denote the sequence Euj by ūj . By the Gagliardo-Nirenberg inequality, if

np
1≤q< n−p ,
sup kukLq (Ω) ≤ sup kūkLq (Rn ) ≤ C sup kūj kW 1,p (Rn ) ≤ CM .
For > 0, let η denote the standard mollifiers and set ūj = η ∗ Euj .
By choosing > 0 sufficiently small, ūj ∈ C0∞ (Ω̃). Since
1 y
Z Z
ūj = η( )ūj (x − y)dy = η(z)ūj (x − z)dz ,
B(0,) n B(0,1)
and if ūj is smooth,

1 1
d
Z Z
ūj (x − z) − ūj (x) = ūj (x − tz)dt = − Dūj (x − tz) · z dt .
0 dt 0
Hence,
Z Z 1
|ūj (x) − ūj (x)| = η(z) |Dūj (x − tz)| dzdt ,
B(0,1) 0
so that
Z Z Z 1Z
|ūj (x) − ūj (x)|dx = η(z) |Dūj (x − tz)| dxdzdt
Ω̃ B(0,1) 0 Ω̃
≤ kDūj kL1 (Ω̃) ≤ kDūj kLp (Ω̃) < CM .
Using the Lp -interpolation Lemma 2.111,
kūj − ūj kLq (Ω̃) ≤ kūj − ūj kaL1 (Ω̃) kūj − ūj k1−a
np
L n−p (Ω̃)
1−a
≤ CM kDūj − Dūj kLp (Ω̃)
≤ CM M 1−a (2.24)
The inequality (2.24) shows that ūj is arbitrarily close to ūj in Lq (Ω)
uniformly in j ∈ N; as such, we attempt to use the smooth sequence ūj to
construct a convergent subsequence ūjk . Our goal is to employ the Arzela-
Ascoli Theorem, so we show that for > 0 fixed,
kūj kC 0 (Ω̃) ≤ M̃ < ∞ and ūj is equicontinous.
For x ∈ Rn ,
Z
sup kūj kC 0 (Ω̃) ≤ sup η (x − y)|ūj (y)|dy
j j B(x,)
≤ kη kL∞ (Rn ) sup kūj kL1 (Ω̃) ≤ C−n < ∞ ,
j
and similarly
sup kD̄uj kC 0 (Ω̃) ≤ kDη kL∞ (Rn ) sup kūj kL1 (Ω̃) ≤ C−n−1 < ∞ .
j j
The latter inequality proves equicontinuity of the sequence ūj , and hence
there exists a subsequence ujk which converges uniformly on Ω̃, so that
lim sup kūjk − ūjl kLq (Ω̃) = 0 .

k,l→∞
It follows from (2.24) and the triangle inequality that
lim sup kūjk − ūjl kLq (Ω̃) ≤ C .

k,l→∞
Letting C = 1, 21 , 13 , etc., and using the diagonal argument to extract

further subsequences, we can arrange to find a subsequence again denoted
by {ūjk } of {ūj } such that
lim sup kūjk − ūjl kLq (Ω̃) = 0 ,

k,l→∞
and hence
lim sup kujk − ujl kLq (Ω) = 0 ,
k,l→∞
The case that n = p = 2 follows from Theorem 2.90.
Theorem 2.112. Let Ω denote an open, bounded, and smooth domain of

Rn , and let u ∈ H 1 (Ω) . Then u is absolutely continuous on almost all
straight lines parallel to the coordinate axes. Moreover, the weak derivatives
of u coincides with the classical derivative of u almost everywhere.

Proof. It suffices to assume that Ω = x ∈ Rn 0 < xi < 1 , 1 ≤ i ≤ n , and

show that
xn
∂u 0
Z
u(x) = (x , t)dt +const ,
∂xn
|0 {z }
≡v(x)
∂u
where the integrand is the weak derivative of u with respect to xn .
∂xn

Let ω = x ∈ Rn−1 0 < xi < 1 , 1 ≤ i ≤ n − 1 so that Ω = ω × (0, 1) ,

and let ζ ∈ C0∞ (ω) and ϕ ∈ C0∞ (0, 1) be test functions. Since v is absolutely
continuous in xn , integration by parts implies
Z 1 Z 1
0 0
v(x , t)ϕ (t)dt = − vxn (x0 , t)ϕ(t)dt ,
0 0
where vxn denotes the classical derivative of v with respect to xn . Multi-

plying both sides by ζ(x0 ) and integrating over ω , we find that
Z Z
v(x)ζ(x )ϕ (xn )dx = − vxn (x)ζ(x0 )ϕ(xn )dx .
0 0
Ω Ω
By the definition of weak derivative,

∂u
Z Z
0 0
u(x)ζ(x )ϕ (xn )dx = − (x)ζ(x0 )ϕ(xn )dx .
Ω Ω ∂xn
∂u
Since the classical derivative vxn is the same as , the right-hand side of
∂xn
the two equalities above are the same; hence due to the fact that the test
function ζ ∈ C0∞ (ω) is arbitrary,
Z 1
u(x0 , xn ) − v(x0 , xn ) ϕ0 (xn )dxn = 0

0
for almost all x0 ∈ ω . As a consequence, by Problem 2.16, we find that
u(x0 , xn ) − v(x0 , xn ) = a constant independent of xn
which shows that u is absolutely continuous on almost all straight lines

parallel the xn -axis.
2.2.13 Exercises
Problem 2.12. Suppose that 1 < p < ∞. If τy f (x) = f (x − y), show that
f belongs to W 1,p (Rn ) if and only if τy f is a Lipschitz function of y with
values in Lp (Rn ), i.e.
kτy f − τz f kLp (Rn ) ≤ C|y − z| .
What happens in the case p = 1?

Problem 2.13. If for j = 1, 2 and pj ∈ [1, ∞] and uj ∈ Lpj , show that

u1 u2 ∈ Lr provided that 1/r = 1/p1 + 1/p2 and
ku1 u2 kLr ≤ ku1 kLp1 ku2 kLp2 .
Show that this implies that the generalized Hölder’s inequality, which states
that if for j = 1, ..., m and pj ∈ [1, ∞] with m 1
P
j=1 pj = 1, then
Z
|u1 · · · um | dx ≤ ku1 kLp1 · · · kum kLpm .
Rn
Problem 2.14. Let f ∈ L1 (R), and set

Z x
g(x) = f (y)dy . (*)
−∞
Continuity of g follows from the Dominated Convergence Theorem. Show

that ∂1 g = f .
(Hint. Given φ ∈ C0∞ (R), use (*) to obtain
Z Z Z x
φ0 (x)g(x)dx = φ0 (x)f (y)dydx .
R R −∞
Then write this integral as

x+h
1 1
Z Z Z
lim [φ(x + h) − φ(x)] g(x)dx = − lim f (y)φ(x) dydx .)
h→0 h R h→0 h R x
Problem 2.15. Show that W n,1 (Rn ) ⊂ C(Rn ) ∩ L∞ (Rn ).

R0 R0
(Hint. u(x) = −∞ · · · −∞ ∂1 · · · ∂n u(x + y)dy1 · · · dyn .)
Problem 2.16. If u ∈ W 1,p (Rn ) for some p ∈ [1, ∞) and ∂j u = 0 on a

connected open set Ω ⊂ Rn , for 1 ≤ j ≤ n, show that u is equal a.e. to a
constant on Ω.
(Hint. Approximate u using that
φi ∗ u → u in W 1,p (Rn ) ,
where φi is a sequence of standard mollifiers. As we showed, given > 0,

we can choose i such that
kφi ∗ u − ukW 1,p (Rn ) < .
Show that ∂j (φi ∗ u) = 0 on Ωi ⊂⊂ Ω, where Ωi % Ω as i → ∞.)
More generally, if ∂j u = fj ∈ C(Ω), 1 ≤ j ≤ n, show that u is equal a.e.

to a function in C 1 (Ω).
Problem 2.17. In case n = 1, deduce from Problems 2.14 and 2.16 that, if
u ∈ L1loc (R) and if ∂1 u = f ∈ L1 (R), then
Z x
u(x) = c + f (y)dy , a.e. x ∈ R ,
−∞
for some constant c.
Problem 2.18. Let Ω := B(0, 12 ) ⊂ R2 denote the open ball of radius 1

2.
For x = (x1 , x2 ) ∈ Ω, let
q
u(x1 , x2 ) = x1 x2 log (| log(|x|)|) where |x| = x21 + x22 .
(a) Show that u ∈ C 1 (Ω̄);
∂2u
(b) show that ∂x2j
∈ C(Ω̄) for j = 1, 2, but that u 6∈ C 2 (Ω̄);
(c) show that u ∈ H 2 (Ω).
Problem 2.19. Theorem 2.84 states that for p > n, and y ∈ B(y, r) ,
1− n
|u(x) − u(y)| ≤ Cr p kDukLp (Rn ) ∀u ∈ C 1 (Rn ) . (2.25)
Prove that the inequality (2.25) in fact holds for all u ∈ W 1,p (Rn ); in par-
ticular, show that Du can be taken to be the weak derivative of u.
Problem 2.20. Let η denote the standard mollifier, and for u ∈ H 2 (R3 ),
set u = η ∗ u. Prove that
√
ku − ukL∞ (R3 ) ≤ C kukH 2 (R3 ) ,
and that
ku − ukL∞ (R3 ) ≤ CkukH 3 (R3 ) .
Problem 2.21. Suppose that for n ≥ 2, Ω ⊂ Rn is a smooth, open, and

bounded domain, and let n denote the outward-pointing unit normal vector
to the boundary ∂Ω. Suppose that u ∈ L2 (Ω) and div u ∈ L2 (Ω). Prove that
u · n ∈ H −1/2 (∂Ω) and that

ku · nkH −1/2 (∂Ω) ≤ C kukL2 (Ω) + k div ukL2 (Ω) .
Problem 2.22. Let Ω ⊂ R2 denote an open, bounded, subset with smooth

boundary. Prove the interpolation inequality:
kDuk2L2 (Ω) ≤ CkukL2 (Ω) kD2 ukL2 (Ω) ∀ u ∈ H 2 (Ω) ∩ H01 (Ω) ,
where D2 u denotes the Hessian matrix of u, i.e., the matrix of second partial
∂2u
derivatives ∂xi ∂xj . Use the fact that C ∞ (Ω)∩H01 (Ω) is dense in u ∈ H 2 (Ω)∩
H01 (Ω).
Problem 2.23. Let D := B(0, 1) ⊂ R2 denote the unit disc, and let
h iα
u(x) = − log |x| .
Prove that the weak derivative of u exists for all α ≥ 0.
Problem 2.24. Suppose that {fn }∞ 1

n=1 is a bounded sequence in H (Ω) for
Ω ⊂ R2 bounded. Show that there exists an f ∈ H 1 (Ω) such that for 1 <
p < 2,
fnl Dfnl * f Df weakly in Lp (Ω) .
Problem 2.25. Suppose that uj * u in W 1,1 (0, 1). Show that uj → u a.e.
2.3 The Fourier Transform
The Fourier transform is one of the most powerful and fundamental tools in
linear analysis, converting constant-coefficient linear differential operators
into multiplication by polynomials. In this section, we define the Fourier
transform, first on L1 (Rn ) functions, next (and miraculously) on L2 (Rn )
functions, and finally on the space of tempered distributions.
2.3.1 Fourier transform on L1 (Rn ) and the space S(Rn )
Definition 2.113. For all f ∈ L1 (Rn ) the Fourier transform F is defined

by Z
n
Ff (ξ) = fˆ(ξ) = (2π)− 2 f (x)e−ix·ξ dx .
Rn
By Hölder’s inequality, F : L1 (Rn ) → L∞ (Rn ).
Definition 2.114. The space of Schwartz functions of rapid decay is de-

noted by
S(Rn ) = {u ∈ C ∞ (Rn ) | xβ Dα u ∈ L∞ (Rn ) ∀α, β ∈ Zn+ }.
It is not difficult to show (as it follows from the definition) that
F : S(Rn ) → S(Rn ) ,
and that
ξ α Dξβ fˆ = (−i)|α| (−1)|β| F(Dxα xβ f ) .
The Schwartz space S(Rn ) is also known as the space of rapidly decreasing
functions; thus, after multiplying by any polynomial functions P (x), P (x)Dα u(x) →
0 as x → ∞ for all α ∈ Zn+ . The classical space of test functions D(Rn ) :=
2
C0∞ (Rn ) ⊂ S(Rn ). The prototype element of S(Rn ) is e−|x| which is not
compactly supported, but has rapidly decreasing derivatives.
The reader is encouraged to verify the following basic properties of S(Rn )
which we will denote by S:
1. S is a vector space.
2. S is an algebra under the pointwise product of functions.

2.3. THE FOURIER TRANSFORM 69
3. P (x)u(x) ∈ S for all u ∈ S and all polynomial functions P (x).
4. S is closed under differentiation.
5. S is closed under translations and multiplication by complex exponen-

tials eix·ξ .
6. S ⊂ L1 (Rn ) (since |u(x)| ≤ C(1 + |x|)n+1 for all u ∈ S and (1 +

|x|)−(n+1) dx decays like |x|−2 as |x| → ∞).
Definition 2.115. For all f ∈ L1 (Rn ), we define operator F ∗ by

Z
n
F ∗ f (x) = (2π)− 2 f (ξ)eix·ξ dξ .
Rn
Lemma 2.116. For all u, v ∈ S(Rn ),
(Fu, v)L2 (Rn ) = (u, F ∗ v)L2 (Rn ) .
Recall that the L2 (Rn ) inner-product for complex-valued functions is

R
given by (u, v)L2 (Rn ) = Rn u(x)v(x)dx.
Proof. Since u, v ∈ S(Rn ), by Fubini’s Theorem,

Z Z
−n
(Fu, v)L2 (Rn ) = (2π) 2 u(x)e−ix·ξ dx v(ξ) dξ
n n
ZR ZR
−n
= (2π) 2 u(x)eix·ξ v(ξ) dξ dx
n n
ZR R Z
−n
= (2π) 2 u(x) eix·ξ v(ξ) dξ dx = (u, F ∗ v)L2 (Rn ) ,
Rn Rn
Theorem 2.117. F ∗ ◦ F = Id = F ◦ F ∗ on S(Rn ).
Proof. We first prove that for all f ∈ S(Rn ), F ∗ Ff (x) = f (x).
Z Z
∗ −n iξ·x −iy·ξ
F Ff (x) = (2π) e e f (y)dy dξ
n Rn
ZR Z
= (2π)−n ei(x−y)·ξ f (y) dy dξ .
Rn Rn
By the dominated convergence theorem,

Z Z
2
F ∗ Ff (x) = lim (2π)−n e−|ξ| ei(x−y)·ξ f (y) dy dξ .
→0 Rn Rn
2
For all > 0, the convergence factor e−|ξ| allows us to interchange the
order of integration, so that by Fubini’s theorem,
Z Z
∗ −n −|ξ|2 i(y−x)·ξ
F Ff (x) = lim (2π) f (y) e e dξ dy .
→0 Rn Rn
Define the integral kernel

Z
−n 2 +ix·ξ
p (x) = (2π) e−|ξ| dξ
Rn
Then
Z
∗
F Ff (x) = lim p ∗ f := p (x − y)f (y)dy .
→0 Rn
2 +ix·ξ
Let p(x) = p1 (x) = (2π)−n e−|ξ|
R
Rn dξ. Then
√ 2 +ix·ξ/√
Z
p(x/ ) = (2π)−n e−|ξ| dξ
n
ZR
2 +ix·ξ n n
= (2π)−n e−|ξ| 2 dξ = 2 p (x) .
Rn
We claim that
1
Z
|x|2
− 4
p (x) = n e and that p(x)dx = 1 . (2.26)
(4π) 2 Rn
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
-4 -2 0 2 4
Figure 2.1: As → 0, the sequence of functions p becomes more localized

about the origin.
Given (2.26), then for all f ∈ S(Rn ), p ∗ f → f uniformly as → 0,

which shows that F ∗ F = Id, and similar argument shows that FF ∗ = Id.
(Note that this follows from the proof of Theorem 2.28, since the standard
mollifiers η can be replaced by the sequence p and all assertions of the
theorem continue to hold, for if (2.26) is true, then even though p does not
R
have compact support, B(0,δ)c p (x)dx → 0 as → 0 for all δ > 0.)
Thus, it remains to prove (2.26). It suffices to consider the case = 12 ;
then by definition
Z
|ξ|2
p 1 (x) = (2π)−n eix·ξ e− 2 dξ
2
Rn

|ξ|2
−n/2 − 2
= F (2π) e .
|x|2
In order to prove that p 1 (x) = (2π)−n/2 e− 2 , we must show that with the
2
|x|2
− 2
Gaussian function G(x) = (2π)−n/2 e ,
G(x) = F(G(ξ)) .
By the multiplicative property of the exponential,

2 /2 2 2
e−|ξ| = e−ξ1 /2 · · · e−ξn /2 ,
it suffices to consider the case that n = 1. Then the Gaussian satisfies the
differential equation
d
G(x) + xG(x) = 0 .
dx
Computing the Fourier transform, we see that
d
−i Ĝ(x) − iξ Ĝ(x) = 0 .
dξ
Thus,
ξ2
Ĝ(ξ) = Ce− 2 .
To compute the constant C,
Z
x2 1
−1
C = Ĝ(0) = (2π) e 2 dx = (2π)− 2
R
which follows from the fact that
Z
x2 1
e 2 dx = (2π) 2 . (2.27)
R
To prove (2.27), one can again rely on the multiplication property of the
exponential to observe that
Z x2 Z x2 Z x2 2
1 2 1 +x2
e 2 dx e 2 dx = e 2 dx
R R R2
Z 2π Z ∞
2
= e−2r rdrdθ = 2π .
0 0

It follows from Lemma 2.116 that for all u, v ∈ S(Rn ),
(Fu, Fv)L2 (Rn ) = (u, F ∗ Fv)L2 (Rn ) = (u, v)L2 (Rn ) .
Thus, we have established the Plancheral theorem on S(Rn ).
Theorem 2.118 (Plancheral’s theorem). F : S(Rn ) → S(Rn ) is an iso-

morphism with inverse F ∗ preserving the L2 (Rn ) inner-product.
2.3.2 The topology on S(Rn ) and tempered distributions
An alternative to Definition 2.114 can be stated as follows:

p
Definition 2.119 (The space S(Rn )). Setting hxi = 1 + |x|2 ,
S(Rn ) = {u ∈ C ∞ (Rn ) | hxik |Dα u| ≤ Ck,α ∀k ∈ Z+ } .
The space S(Rn ) has a Fréchet topology determined by seminorms.
Definition 2.120 (Topology on S(Rn )). For k ∈ Z+ , define the semi-norm
pk (u) = sup hxik |Dα u(x)| ,

x∈Rn ,|α|≤k
and the metric on S(Rn )

∞
X pk (u − v)
d(u, v) = 2−k .
1 + pk (u − v)
k=0
The space (S(Rn ), d) is a Fréchet space.
Definition 2.121 (Convergence in S(Rn )). A sequence uj → u in S(Rn ) if

pk (uj − u) → 0 as j → ∞ for all k ∈ Z+ .
Definition 2.122 (Tempered Distributions). A linear map T : S(Rn ) → C

is continuous if there exists some k ∈ Z+ and constant C such that
|hT, ui| ≤ Cpk (u) ∀u ∈ S(Rn ) .
The space of continuous linear functionals on S(Rn ) is denoted by S 0 (Rn ).

Elements of S 0 (Rn ) are called tempered distributions.
Definition 2.123 (Convergence in S 0 (Rn )). A sequence Tj * T in S 0 (Rn )

if hTj , ui → hT, ui for all u ∈ S(Rn ).
For 1 ≤ p ≤ ∞, there is a natural injection of Lp (Rn ) into S 0 (Rn ) given

by Z
hf, ui = f (x)u(x)dx ∀u ∈ S(Rn ) .
Rn
Any finite measure on Rn provides an element of S 0 (Rn ). The basic example
of such a finite measure is the Dirac delta ‘function’ defined as follows:
hδ, ui = u(0) or, more generally, hδx , ui = u(x) ∀u ∈ S(Rn ) .
Definition 2.124. The distributional derivative D : S 0 (Rn ) → S(Rn ) is

defined by the relation
hDT, ui = −hT, Dui ∀u ∈ S(Rn ) .
More generally, the αth distributional derivative exists in S 0 (Rn ) and is de-
fined by
hDα T, ui = (−1)|α| hT, Dα ui ∀u ∈ S(Rn ) .
Multiplication by f ∈ S(Rn ) preserves S 0 (Rn ); in particular, if T ∈

S 0 (Rn ), then f T ∈ S 0 (Rn ) and is defined by
hf T, ui = hT, f ui ∀u ∈ S(Rn ) .
Example 2.125. Let H := 1[0,∞) denote the Heavyside function. Then

dH
= δ in S 0 (Rn ) .
dx
This follows since for all u ∈ S(Rn ),
∞
dH du du
Z
h , ui = −hH, i = − dx = u(0) = hδ, ui .
dx dx 0 dx
Example 2.126 (Distributional derivative of Dirac measure).
dδ du
h , ui = − (0) ∀u ∈ S(Rn ) .
dx dx
2.3.3 Fourier transform on S 0 (Rn )
Definition 2.127. Define F : S 0 (Rn ) → S 0 (Rn ) by
hFT, ui = hT, Fui ∀u ∈ S(Rn ) ,
with the analogous definition for F ∗ : S 0 (Rn ) → S 0 (Rn ).

Theorem 2.128. FF ∗ = Id = F ∗ F on S 0 (Rn ) .
Proof. By Definition 2.127, for all u ∈ S(Rn )
hFF ∗ T, ui = hF ∗ w, Fui = hT, F ∗ Fui = hT, ui ,
the last equality following from Theorem 2.117.

n
Example 2.129 (Fourier transform of δ). We claim that Fδ = (2π)− 2 .
According to Definition 2.127, for all u ∈ S(Rn ),
Z
n
hFδ, ui = hδ, Fui = Fu(0) = (2π)− 2 u(x)dx ,
Rn
n
so that Fδ = (2π)− 2 .
n
Example 2.130. The same argument shows that F ∗ δ = (2π)− 2 so that
n n
F ∗ [(2π) 2 ] = 1. Using Theorem 2.128, we see that F(1) = (2π)− 2 δ. This
demonstrates nicely the identity
|ξ α û(ξ)| = |F(Dα u)(ξ)|.
In other the words, the smoother the function x 7→ u(x) is, the faster ξ 7→
û(ξ) must decay.
2.3.4 The Fourier transform on L2 (Rn )
In Theorem 2.28, we proved that C0∞ (Rn ) is dense in Lp (Rn ) for 1 ≤ p < ∞.
Since C0∞ (Rn ) ⊂ S(Rn ), it follows that S(Rn ) is dense in Lp (Rn ) as well.
Thus, for every u ∈ L2 (Rn ), there exists a sequence uj ∈ S(Rn ) such that
uj → u in L2 (Rn ), so that by Plancheral’s Theorem 2.118,
kûj − ûk kL2 (Rn ) = kuj − uk kL2 (Rn ) < .
It follows from the completeness of L2 (Rn ) that the sequence ûj converges
in L2 (Rn ).
Definition 2.131 (Fourier transform on L2 (Rn )). For u ∈ L2 (Rn ) let uj

denote an approximating sequence in S(Rn ). Define the Fourier transform
as follows:
Fu = û = lim ûj .
j→∞
Note well that F on L2 (Rn ) is well-defined, as the limit is independent

of the approximating sequence. In particular,
kûkL2 (Rn ) = lim kûj kL2 (Rn ) = lim kuj kL2 (Rn ) = kukL2 (Rn ) .
j→∞ j→∞
By the polarization identity
1
(u, v)L2 (Rn ) = ku + vk2L2 (Rn ) − iku + ivk2L2 (Rn ) − (1 − i)kuk2L2 (Rn ) − (1 − i)kvk2L2 (Rn )
2
we have proved the Plancheral theorem1 on L2 (Rn ):
Theorem 2.132. (u, v)L2 (Rn ) = (Fu, Fv)L2 (Rn ) ∀u, v ∈ L2 (Rn ) .
2.3.5 Bounds for the Fourier transform on Lp (Rn )

n
We have shown that for u ∈ L1 (Rn ), kûkL∞ (Rn ) ≤ (2π)− 2 kukL1 (Rn ) , and
that for u ∈ L2 (Rn ), kûkL2 (Rn ) = kukL2 (Rn ) . Interpolating p between 1 and
2 yields the following result.
Theorem 2.133 (Hausdorff-Young inequality). If u ∈ Lp (Rn ) for 1 ≤ p ≤

p−1
2, then for q = p , there exists a constant C such that
kûkLq (Rn ) ≤ CkukLp (Rn ) .
Returning to the case that u ∈ L1 (Rn ), not only is Fu ∈ L∞ (Rn ), but

the transformed function decays at infinity.
Theorem 2.134 (Riemann-Lebesgue “lemma”). For u ∈ L1 (Rn ), Fu is

continuous and Fu(ξ) → 0 as |ξ| → ∞.
Proof. Let BM = B(0, M ) ⊂ Rn . Since f ∈ L1 (Rn ), for each > 0, we can

choose M sufficiently large such that fˆ(ξ) ≤ + BM e−ix·ξ |f (x)|dx. Using
R
Lemma 2.23, choose a sequence of simple functions φj (x) → f (x) a.e. on

BM . For jnN chosen sufficiently large,
Z
ˆ
f (ξ) ≤ 2 + φj (x)e−ix·ξ dx .
BM
1
The unitarity of the Fourier transform is often called Parseval’s theorem in science
and engineering fields, based on an earlier (but less general) result that was used to prove
the unitarity of the Fourier series.
PN
Write φj (x) = l=1 Cl 1El (x) so that
N Z
fˆ(ξ) ≤ 2 +
X
Cl φj (x)e−ix·ξ dx .
l=1 El
By the regularity of the Lebesgue measure µ, for all > 0 and each l ∈
{1, ..., N }, there exists a compact set Kl and an open set Ol such that
µ(Ol ) − /2 < µ(El ) < µ(Kl ) + /2 .
Then Ol = {∪α∈Al Vαl | Vlα ⊂ Rn is open rectangle , Al arbitrary set }, and

Kl ⊂ ∪N l
j=1 Vj ⊂ Ol where {1, ..., Nl } ⊂ Al such that
l
|µ(El ) − µ(∪N l
j=1 Vj )| < .
l
It follows that Z Z

e−ix·ξ dx − N e−ix·ξ dx < .

El l
∪j=1 Vjl
On the other hand, for each rectangle Vjl , V l e−ix·ξ dx| ≤ C/(ξ1 · · · ξn ), so
R
j
that
1
fˆ(ξ) ≤ C + .
ξ1 · · · ξn
Since > 0 is arbitrary, we see that fˆ(ξ) → 0 as |ξ| → ∞. Continuity of Fu
follows easily from the dominated convergence theorem.
2.3.6 Convolution and the Fourier transform
Theorem 2.135. If u, v ∈ L1 (Rn ), then u ∗ v ∈ L1 (Rn ) and

n
F(u ∗ v) = (2π) 2 Fu Fv .
Proof. Young’s inequality (Theorem 2.53) shows that u ∗ v ∈ L1 (Rn ) so

that the Fourier transform is well-defined. The assertion then follows from
a direct computation:
Z
n
F(u ∗ v) = (2π)− 2 e−ix·ξ (u ∗ v)(x)dx
Rn
Z Z
n
= (2π)− 2 u(x − y)v(y)dy e−ix·ξ dx
n n
ZR ZR
n
= (2π)− 2 u(x − y)e−i(x−y)·ξ dx v(x) e−iy·ξ dy
Rn Rn
n
= (2π) ûv̂ (by Fubini’s theorem) .
2

By using Young’s inequality (Theorem 2.54) together with the Hausdorff-

Young inequality, we can generalize the convolution result to the following
Theorem 2.136. Suppose that u ∈ Lp (Rn ) and v ∈ Lq (Rn ), and let r

r
1 1 1
satisfy r = p + q − 1 for 1 ≤ p, q, r ≤ 2. Then F(u ∗ v) ∈ L r−1 (Rn ) and
n
F(u ∗ v) = (2π) 2 Fu Fv .
2.3.7 An explicit computation with the Fourier Transform
The computation of the Green’s function for the Laplace operator is an

important application of the Fourier transform. For this purpose, we will
compute fˆ for the following two cases: (1) f (x) = e−t|x| , t > 0 and (2)
f (x) = |x|α , −n < α < 0.
Case (1) In this case, f is rapidly decreasing but not in the Schwartz class
S(Rn ). We begin with n = 1. It follows that
Z ∞ Z 0 Z ∞
−t|x| −t|x| −ix·ξ x(t−iξ)
[F(e )](ξ) = e e dµ1 (x) = e dµ1 (x) + ex(−t−iξ) dµ1 (x)
−∞ −∞ 0
r
1 e
h x(t−iξ) 0 e x(−t−iξ) ∞ i 2 t
=√ + = .

2π t − iξ −∞ −t − iξ 0 π t + ξ2
2

1 ∞ t
Z
−t|x|
By the inversion formula, we then see that e = eixξ dξ.
π −∞ t + ξ 2
2
Next, when n > 1 we will show that for some function g(t, s),
Z ∞
−t|x| 2
e = g(t, s)e−s|x| ds . (2.28)
0
In order to determine g(t, s), we suppose that (2.28) holds, and compute its
Fourier transform:
Z ∞ Z ∞ 1 n −|ξ|2
−t|x| −s|x|2
F(e )= g(t, s)F(e )ds = g(t, s) √ e 4s ds ,
0 0 2s
where we have used the definition of the Fourier transform of the Gaussian
function given in the proof of Theorem 2.117. We are thus seeking a function
g(t, s) which satisfies
Z ∞
−tλ 2
e = g(t, s)e−sλ ds, ∀ λ > 0.
0
We begin by computing
Z ∞ 2 2
−st2 −sξ 2 e−s(t +ξ ) ∞ 1
e e ds = 2 2 = 2 . (2.29)
0 −(t + ξ ) 0 t + ξ2
With λ = |x| > 0, we use (2.29) to find that

1 ∞ t 1 ∞ ∞ −st2 −sξ2 iλξ
Z Z Z
−tλ iλξ
e = e dξ = t e e ds e dξ
π −∞ t2 + ξ 2 π −∞ 0
1 ∞ ∞ −sξ2 iλξ −st2
Z ∞
t
Z Z
2 |x|2
= t e e dξ e ds = √ e−st e− 4s ds
π 0 −∞ 0 sπ
so that
√ n t 2
g(t, s) = 2s √ e−st ,
sπ
and hence
∞ ∞
t
Z Z
|x|2n 2 2
−t|x| −
F(e )(ξ) = g(t, s)F(e √ (2s) 2 e−s(t +|ξ| ) ds
4s )ds =
0 0 sπ
∞
t 1 C(n)t
Z
n
= n+1 √ (2s) 2 e−s ds = n+1 ,
(t2 + |ξ|2 ) 2 0 πs (t2 + |ξ|2 ) 2
Z ∞ r
1 n 2n n + 1
where the constant C(n) = √ (2s) 2 ds = Γ( ), and Γ is the
0 πs π 2
so-called gamma-function. It follows that
r
2n n + 1 t
F −1 (e−t|ξ| )(x) = F(e−t|ξ| )(−x)= Γ( ) n+1 . (2.30)
π 2 (t + |x|2 ) 2
2
Case (2) For this case, we compute F(|x|α ), when −n < α < 0. Using the
definition of Γ(n) above, we see that
Z ∞ ∞
α
Z
−α −1 −s|x| 2 α
s 2 e ds = |x| α
s− 2 −1 e−s ds = |x|α Γ(− ) ,
0 0 2
Therefore,
∞ ∞
1 1
Z Z
|ξ|2
−α −1 −s|x|2 α n
α
F(|x| ) = s 2 F(e )ds = n s− 2 − 2 −1 e− 4s ds
Γ(− α2 ) 0 2 Γ(− α2 )
2 0
n
1 ∞
|ξ|2 − α2 − n2 −1 −s |ξ|2 2α+ 2 Γ( α+n
2 )
Z
= n e 2
ds = α |ξ|−α−n ,
2 2 Γ(− α2 ) 0 4s 4s Γ(− 2 )
where we impose the condition −n < α < 0 to ensure the boundedness of

the Γ-function. In particular, for n = 3 and α = −1,
√ r
−1 2Γ(1) −2 2 −2
F(|x| ) = 1 |ξ| = |ξ| ,
Γ( 2 ) π

r
−1 −2 π 1
F (|ξ| )= . (2.31)
2 |x|
2.3.8 Applications to the Poisson, Heat, and Wave equations

The Poisson equation on R3
In Theorem 2.105, we proved the existence of unique weak solutions to the

Dirichlet problem on a bounded domain Ω. We will now provide an explicit
representation for solutions to the Poisson problem on R3 . The issue of
uniqueness in this setting will be of interest.
Given the Poisson problem
∆u = f in S 0 ,
we compute the Fourier transform of both sides to obtain that
−|ξ|2 û(ξ) = fˆ(ξ) . (2.32)
Distributional solutions to (2.32) are not unique; for example,
fˆ(ξ) fˆ(ξ)
û(ξ) = − and û(ξ) = − +δ
|ξ|2 |ξ|2
are both solutions. By requiring solutions to have enough decay, such as
u ∈ L2 (Rn ) so that û ∈ L2 (Rn ), then we do obtain uniqueness.
We will find an explicit representation for the solution to the Poisson
problem when n = 3. If u ∈ L2 (R3 ), then using (2.31), we see that
fˆ(ξ) fˆ(ξ)
û(ξ) = − 2 ⇒ u(x) = −F −1 (x)
|ξ| |ξ|2
h i
= − F −1 (|ξ|−2 ) ? F −1 (fˆ) (x) = (Φ ∗ f )(x) ,
1
where Φ(x) = − . The function Φ is the so-called fundamental solution;
4π|x|
more precisely, it is the distributional solution of the equation
∆Φ = δ in S 0 .
Conceptually
∆(Φ ∗ f ) = ∆Φ ∗ f = δ∗f = f ∀ f ∈ C(Rn ) whenever Φ ∗ f makes sense ,
where the first equality follows from the fact that

n n
h∆(Φ ∗ f ), ϕ̂i = (2π) 2 h−|ξ|2 Φ̂fˆ, ϕi = (2π) 2 hF(∆Φ), fˆϕi
n
= (2π) 2 h∆Φ, F(fˆϕ)i = h∆Φ, fe ∗ ϕ̂i = h∆Φ ∗ f, ϕ̂i .
Example 2.137. On R2 , ∆(ex1 cos x2 ) = 0. The function ex1 cos x2 is not

a tempered distribution because it grows too fast as x1 → ∞. As such, the
Fourier transfor of ex1 cos x2 is not defined.
Using Fourier transform to convert PDE to linear algebraic equations
only provides those solutions which do not grow too rapidly at ∞.
The Poisson integral formula on the half-space
Let Ω = Rn × R+ =, and consider the Dirichlet problem

h ∂2 ∂2 ∂2 i h ∂2 i
+ + · · · + u = + ∆ u = 0 in Ω × (0, ∞) ,
∂t2 ∂x21 ∂x2n ∂t2
u(x, 0) = f (x)(∈ S(Rn )) on ∂Ω × (0, ∞) .
(Note that for any constant c, c t is always a solution as it is harmonic

and vanishes at the boundary t = 0.) For uniqueness, we insist that u be
bounded. This in turn means u is in S 0 and hence we may use the Fourier
transform. Applying the Fourier transform (in the x variable) Fx , we see
that
∂2
Fx u(ξ, t) − |ξ|2 Fx u(ξ, t) = 0 , Fx u(ξ, 0) = fˆ(ξ) .
∂t2
Therefore, Fx u(ξ, t) = C1 (ξ)et|ξ| + C2 (ξ)e−t|ξ| , and C1 (ξ) = 0 by the growth
condition imposed on u. Then Fx u(ξ, t) = fˆ(ξ)e−t|ξ| and hence using (2.30),
h i
u(x, t) = F −1 (fˆ(ξ)e−t|ξ| )(x) = F −1 (e−t|ξ| ) ∗ f (x)
Γ( n+1
2 ) tf (y)
Z
= n+1 n+1 dy .
π 2 Rn (t2 + |x − y|2 ) 2
This is the Poisson integral formula on the half-space.

If f is bounded, i.e., f ∈ L∞ (Rn ), then the integral converges and u ∈
L∞ (Rn × R+ ). Therefore, u ∈ C ∞ (Rn × R+ ) ∩ L∞ (Rn × R+ ).
The Heat equation
Let t ≥ 0 denote time, and x denote a point in space Rn . The function u(x, t)
denotes the temperature at time t and position x, and g ∈ S(Rn ) denotes
the initial temperature distribution. We wish to solve the heat equation
ut (x, t) = ∆u(x, t) in Rn × (0, ∞) , (2.33a)

u(x, 0) = g(x) on Rn × {t = 0} . (2.33b)
Taking the Fourier transform of (2.33), we find that
∂t û(ξ, t) = −|ξ|2 û(ξ, t) ,

û(ξ, 0) = ĝ(ξ) .
2t
Therefore, û(ξ, t) = ĝ(ξ)e−|ξ| and hence
2
h 2
i
u(x, t) = F −1 ĝ(ξ)e−|ξ| t (x) = F −1 e−|ξ| t ∗ g (x)
1
Z
|x−y|2
− 4t
= e g(y)dy (≡ (H(·, t) ∗ g)(x)) . (2.34)
(4πt)n/2 Rn
Theorem 2.138. If g ∈ L∞ (Rn ), then the solution u to (2.33) is in C ∞ (Rn ×

(0, ∞)).
2
e−|x| /4t
Proof. The function is C ∞ (Rn × [α, ∞)) for all α > 0.
(4πt)n/2
Remark 2.139. The representation formula (2.34) shows that whenever g

is bounded, continuous, and positive, the solution u(x, t) to (2.33) is positive
everywhere for t > 0.
The representation formula (2.34) can also be used to prove the following
Theorem 2.140. Assume that g ∈ C(Rn ) ∩ L∞ (Rn ). Then u defined by

(2.34) is continuous at t = 0, that is,
lim u(x, t) = g(x0 ) ∀ x0 ∈ Rn .

(x,t)→(x0 ,0+ )
In order to study the Inhomogeneous heat equation
ut (x, t) − ∆u(x, t) = f (x, t) in Rn × (0, ∞) , (2.35a)

u(x, 0) = 0 on Rn × {t = 0} , (2.35b)
we introduce the parameter s > 0, and consider the following problem for
U:
Ut (x, t, s) = ∆U (x, t, s) ,
U (x, s, s) = f (x, s) .
Then by (2.34),
Z
U (x, t, s) = H(x − y, t − s)f (y, s)dy .
Rn
We next invoke Duhamel’s principle to find a solution u(x, t) to (2.35):

Z t Z tZ
u(x, t) = U (x, t, s)ds = H(x − y, t − s)f (y, s)dyds . (2.36)
0 0 Rn
The principle of linear superposition then shows that the solution of the
problem
ut (x, t) − ∆u(x, t) = f (x, t) in Rn × (0, ∞) ,

u(x, 0) = g(x) on Rn × {t = 0} ,
is the sum of (2.34) and (2.36):

Z tZ Z
u(x, t) = H(x − y, t − s)f (y, s)dyds + H(x − y, t)g(y)dy
0 Rn Rn
Z t
= [H(·, t) ∗ g](x) + [H(·, t − s) ∗ f (·, s)](x)ds . (2.37)
0
The Wave equation
For wave speed c > 0, and for x ∈ Rn , t ∈ R, consider the following second-
order linear hyperbolic equation:
utt (x, t) = c2 ∆u(x, t) in Rn × (0, ∞) ,

u(x, 0) = f (x) on Rn × {t = 0} ,
ut (x, 0) = g(x) on Rn × {t = 0} .
Taking the Fourier transform of (2.38), we find that
ûtt (ξ, t) = −c2 |ξ|2 û(ξ, t) in Rn × (0, ∞) ,

û(ξ, 0) = fˆ(ξ) on Rn × {t = 0} ,
ût (ξ, 0) = ĝ(ξ) on Rn × {t = 0} .
The general solution of this second-order ordinary differential equations is

given by
û(ξ, t) = C1 (ξ) cos c|ξ|t + C2 (ξ) sin c|ξ|t .
Solving for C1 and C2 by using the initial conditions, we find that
sin c|ξ|t
û(ξ, t) = fˆ(ξ) cos c|ξ|t + ĝ(ξ) .
c|ξ|
Therefore,
h sin c|ξ|t i
u(x, t) = F −1 (cos c|ξ|t) ∗ f + F −1 ∗ g (x)
c|ξ|
1 h d −1 sin c|ξ|t sin c|ξ|t i
= F ∗ f + F −1 ∗ g (x) .
c dt |ξ| |ξ|
For the case that n = 1,
m
sin ctλ −ixλ m
ei(ct−x)λ − e−i(ct+x)λ
Z Z
e dλ = dλ .
−m λ −m 2iλ
m
eiz
Z
By the Cauchy integral formula and the residue theorem, lim dz = iπ .
m→∞ −m z
Therefore, ∀ t > 0,
m
1 sin ctλ −ixλ
Z
1 |x| < ct
lim e dλ = χ|x|<ct (x) = .
m→∞ π −m λ 0 |x| ≥ ct
sin c|ξ|t r
−1 π
Corollary 2.141. F (x) = χ (x) in S 0 (R).
|ξ| 2 |x|<ct
Proof. For all ϕ ∈ S(R),

Z sin c|ξ|t Z
sin c|ξ|t −1
−1
F (x)ϕ(x)dx = F (ϕ)(ξ)dξ
R |ξ| R |ξ|
Z mZ
1 sin c|ξ|t ixξ
= lim √ e ϕ(x)dxdξ ,
m→∞ 2π −m R |ξ|
and by Fubini’s theorem together with the dominated convergence theorem,

we see that
m Z Z m
sin c|ξ|t ixξ sin c|ξ|t ixξ
Z Z
lim e ϕ(x)dxdξ = lim e ϕ(x)dξdx
m→∞ −m R |ξ| m→∞ R −m |ξ|
Zˇ Z m
sin c|ξ|t ixξ
Z
= lim e ϕ(x)dξdx = π χ|x|<ct (x)ϕ(x)dx .
R m→∞ −m |ξ| R

We have thus established d’Alembert’s formula for the solution of the

the 1-D wave equation:
r
1 π 1 hd
Z Z i
u(x, t) = √ f (x − y)χ|y|<ct (y)dy + g(x − y)χ|y|<ct (y)dy
c 2 2π dt R R
Z ct Z ct
1 d 1
= f (x − y)dy + g(x − y)dy
2c dt −ct 2c −ct
f (x − ct) + f (x + ct) 1 x+ct
Z
= + g(y)dy .
2 2c x−ct
We have just used the Fourier transform to find explicit solutions to

the fundamental linear elliptic, parabolic, and hyperbolic equations. More
generally, the Fourier transform is a powerful tool for the analysis of many
other constant coefficient linear partial differential equations.
2.3.9 Exercises
Problem 2.26. The general solution u(x, t) to the 1-D wave equation
utt (x, t) = c2 uxx (x, t) in R × (0, ∞)
can be expressed as u(x, t) = F (x + ct) + G(x − ct) for some functions F

and G. Use this expression to solve the initial-value problem
utt (x, t) = c2 uxx (x, t) in R × (0, ∞) ,

u(x, 0) = f (x) on R × {t = 0} ,
ut (x, 0) = g(x) on R × {t = 0} .
Check if your answer agrees with the d’Alembert formula.
Problem 2.27. (a) Using the Euler identity, compute F −1 (cos c|ξ|t) di-
sin c|ξ|t
rectly for the case n = 1. Find F −1 via the formula
|ξ|
1 d −1 sin c|ξ|t
F −1 (cos c|ξ|t) = F .
c dt |ξ|
(b) For the case that n = 3, compute
−1 sin c|ξ|t sin c|ξ|t ix·ξ

Z
F (x) = lim e dµ3 (ξ) .
|ξ| m→∞ |ξ|≤m |ξ|
Using spherical coordinates, show that
1 3 Z m Z π Z 2π sin cρt
−1 sin c|ξ|t

F (x) = lim √ ei|x|ρ cos φ ρ2 sin φdθdφdρ
|ξ| m→∞ 2π ρ
Z m 0 0 0
1 sin cρt i|x|ρ cos φ φ=π
=√ lim e dρ
2π m→∞ 0 −i|x|

φ=0
Z m −icρt
1 e − eicρt h i|x|ρ i
=√ lim e − e−i|x|ρ dρ
2π m→∞ 0 2|x|
Z ∞h
1 i
= eiρ(|x|−ct) − eiρ(|x|+ct) dµ1 (ρ) .
2|x| −∞
(c) For t > 0 ,

h sin c|ξ|t i
F −1 ? ϕ (x)
|ξ|
Z ∞Z ∞
1 1 iρ(r−ct)
Z
= e ϕ(x − rω)r2 dµ1 (ρ)dµ1 (r)dSω
4π ∂B(0,1) 0 −∞ r
ct 1
Z Z
= ϕ(x − ctw)dSω = ϕ(y)dSy ,
4π ∂B(0,1) 4πct ∂B(x,ct)
and conclude that
1 ∂h 1
Z Z i
u(x, t) = g(y)dSy + f (y)dSy .
4πc2 t ∂B(x,ct) ∂t 4πc2 t ∂B(x,ct)
(2.39)
Problem 2.28. Derive the solution to the 2-D and 3-D wave equations
utt (x, t) = c2 ∆u(x, t) in Rn × (0, ∞) , (2.40a)

u(x, 0) = f (x) on Rn × {t = 0} , (2.40b)
ut (x, 0) = g(x) on Rn × {t = 0} , (2.40c)
by the method of spherical means.

Let Mh (x, r) denote the average of the function h on a sphere with center
x and radius r (hence x and r are independent variable here), that is,
1
Z
Mh (x, r) = h(y)dSy . (2.41)
ωn−1 rn−1 ∂B(x,r)
Setting y = x + rω with |ω| = 1 , we get

1
Z
Mh (x, r) = h(x + rω)dSω . (2.42)
ωn−1 ∂B(0,1)
Note that in (2.41) Mh (x, r) is defined only for r > 0 , while we can extend
its definition to all r ∈ R using (2.42). Therefore, we treat Mh (x, r) as a
function on Rn+1 . If h ∈ C k (Rn ) , then Mh ∈ C k (Rn+1 ) . For h ∈ C 2 (Rn ) ,
n
∂ 1
Z X
Mh (x, r) = hxi (x + rω)ωi dSω
∂r ωn−1 ∂B(0,1) i=1
r
Z
by divergence theorem = ∆x h(x + ry)dy
ωn−1 B(0,1)
Z rZ
r1−n
= ∆x h(x + ρω)ρn−1 dSω dρ
ωn−1 0 ∂B(0,1)
Z r
1−n
=r ∆x ρn−1 Mh (x, ρ)dρ .
0
Therefore, for all h ∈ C 2 (Rn ) ,

∂ h n−1 ∂ i
r Mh (x, r) = rn−1 ∆x Mh (x, r)
∂r ∂r
h ∂2 n−1 ∂ i
⇒ + Mh (x, r) = ∆x Mh (x, r) .
∂r2 r ∂r
The last equality is called the Darboux equation. By (2.42), Mh (x, r) is even
∂
in r , so Mh (x, 0) = h(x) and Mh (x, r) = 0.

∂r r=0
Now let u be a solution to (2.40), and let Mu (x, r, t) be the spherical

mean of u , that is,
1
Z
Mu (x, r, t) = u(x + rω, t)dSω .
ωn−1 ∂B(0,1)
Then Mu (x, 0, r) = u(x, t) . Moreover,

1
Z
∆x M u = ∆x u(x + rω, t)dSω
ωn−1 ∂B(0,1)
1 ∂2 h 1 1 ∂2
Z i
= 2 2 u(x + rω, t)dSω = 2 2 Mu .
c ∂t ωn−1 ∂B(0,1) c ∂t
By the Darboux equation,
∂2 h 2
2 ∂ n−1 ∂ i
M u = c + Mu . (2.43)
∂t2 ∂r2 r ∂r
We then transform a n + 1-dimensional PDE to a 1 + 1-dimensional PDE.

n = 3: Let n = 3 in (2.43), we find that
∂2 2 ∂
2
(rM u ) = c (rMu ) .
∂t2 ∂r2
By the d’Alembert formula (here r plays the role of x in the d’Alembert
formula), we find that
1h i
Mu (x, r, t) = (r + ct)Mf (x, r + ct) + (r − ct)Mf (x, r − ct)
2r
Z r+ct
1
+ sMg (x, s)ds .
2cr r−ct
Use lim Mu (x, r, t) = u(x, t) to find u . Does this agree with (2.39)?
r→0
Problem 2.29. By (2.39), the solution of the 3-dimensional wave equation
utt (x, t) = c2 ∆u(x, t) in R3 × (0, ∞) , (2.44a)

u(x, 0) = f (x) on R3 × {t = 0} , (2.44b)
ut (x, 0) = g(x) on R3 × {t = 0} , (2.44c)
can be expressed as
1 ∂h 1
Z Z i
u(x, t) = g(y)dS y + f (y)dSy .
4πc2 t ∂B(x,ct) ∂t 4πc2 t ∂B(x,ct)
Suppose that f ∈ C02 (R3 ) and g ∈ C01 (R3 ) so that they provide a solution
u ∈ C 2 (R3 × (0, ∞)) . Show that there exists a constant K > 0 so that
K
|u(x, t)| ≤ ∀t > 0.
t
Hint: First rewrite (2.39) as
3
1
Z h X i
u(x, t) = 2 2 2 tg(y) + f (y) + fyi (y)(yi − xi ) dSy
4π c t |y−x|=ct
i=1
and convert the integral into an integral over B(x, ct) .
Proof. By the change of variable y = x + ctω ,

∂ ∂
Z Z
f (y)dSy = f (x + ctω)c2 t2 dSω
∂t ∂B(x,ct) ∂t ∂B(0,1)
Z h i
= 2c2 tf (x + ctω) + (Df )(x + ctω)c3 t2 ωi dSω
∂B(0,1)
1h
Z i
= 2f (y) + (Df )(y) · (y − x) dSy ,
∂B(x,ct) t
we find that
3
1
Z h X i
u(x, t) = 2 2 2 tg(y) + f (y) + fyi (y)(yi − xi ) dSy
4π c t |y−x|=ct i=1
3 Z
1 X h tg(y) + f (y) i
= (yi − xi ) + ctfyi (y) Ni dSy
4π 2 c2 t2 ∂B(x,ct) ct
i=1
1
Z h 3tg(y) + 3f (y) + t(Dg)(y) + (Df )(y) · (y − x) i
= 2 2 2 + ct(∆f )(y) dy ;
4π c t B(x,ct) ct
hence
C |f (y)|
Z h i
|u(x, t)| ≤ |g(y)| + t|Dg(y)| + + |Df (y)| + t|∆f (y)| dy
t2 B(x,ct) t
Ch i
≤ kgk 23 3 + kDgkL1 (R3 ) + kf kL3 (R3 ) + kDf k 23 3 + k∆f kL1 (R3 ) .
t L (R ) L (R )
Problem 2.30. Let us consider the BBM equation
ut + ux + uux − uxxt = 0 ∀ x ∈ R , t ∈ (0, T ] , (2.45a)

u(x, 0) = g(x) ∀x ∈ R. (2.45b)
(1) Use the Fourier transform to show that a bounded solution to (2.45)
satisfies
Z tZ ∞ h 1 i
u(x, t) = g(x) + K(x − y) u(y, s) + u2 (y, s) dyds , (2.46)
0 −∞ 2
where K is defined by
1
K(x) = sgn(x)e−|x| .
2
(2) Write (2.46) as u = F (u) , that is, treat the right-hand side of (2.46)
as a function of u . Show that for T > 0 small enough, F has a fixed-
point in the space of bounded continuous functions. (Hint: similar to
the proof of the fundamental theorem of ODE, you can try to show that
the map F is a contraction mapping if T is small enough, and then
apply the contraction mapping theorem.)
1 RR
Problem 2.31. (a) For f ∈ L1 (R), set SR f (x) = (2π)− 2 ˆ ixξ dξ.
−R f (ξ)e
Show that Z ∞
SR f (x) = KR ∗ f (x) = KR (x − y)f (y)dy
−∞
where
R
sin Rx
Z
KR (x) = (2π)−1 eixξ dξ = .
−R πx
(b) Show that if f ∈ L2 (R), then SR f → f in L2 (R) as R → ∞.
Problem 2.32. Show that for any R ∈ (0, ∞), there exists f ∈ L1 (R) such
that SR f 6∈ L1 (R). (Hint. Note that KR 6∈ L1 (R).) For partial credit,
explain why the result is interesting.
Problem 2.33. Assume w ∈ S 0 ∩ L1loc (Rn ) and w(x) ≥ 0. Show that if

ŵ ∈ L∞ (Rn ), then w ∈ L1 (Rn ) and
kŵkL∞ (Rn ) = (2π)−n/2 kwkL1 (Rn )
(Hint. Consider wj (x) = ψ( xj )w(x) with ψ ∈ C0∞ (Rn ) and ψ(0) = 1. Use
the fact that wj * w in S 0 .)
Problem 2.34. Consider the Poisson equation on R1 : uxx = f .

x + |x| |x|
(a) Show that ϕ(x) = and ψ(x) = are both distributional solu-
2 2
tions to uxx = δ0 .
(b) Let f be continuous with compact support in R . Show that

Z
u(x) = ϕ(x − y)f (y)dy
R
and Z
v(x) = ψ(x − y)f (y)dy
R
both solve the Poisson equation wxx (x) = f (x) (without relying upon
distribution theory).
Problem 2.35. Let T ∈ S 0 (Rn ) and f ∈ S(Rn ) . Show that the Leibniz rule
∂ ∂T
for distributional derivatives holds; that is, show that (f T ) = f +
∂xi ∂xi
∂f
T in the sense of distribution.
∂xi
2 2
Problem 2.36. Let f (x) = e−s|x| and g(x) = e−t|x| . Find the Fourier
transform of f (and g) and use the inversion formula to compute f ∗ g .
Problem 2.37. Let dr denote the map given by dr f (x) = f (rx) . Show that
F(dr f ) = r−n d1/r F(f ) .
Problem 2.38. Show that a function f ∈ L2 (Rn ) is real if and only if

fˆ(−ξ) = fˆ(ξ) .
2
Problem 2.39. Find the Fourier transform of the function f (x) = xetx
for t < 0 .
Problem 2.40. Find the Fourier transform of 1(−a,a) , the characteristic

(indicator) function of the set (−a, a) .
Problem 2.41. Let f (x) = 1(0,∞) (x)e−tx , that is,
e−tx if x > 0 ,

f (x) =
0 if x ≤ 0 .
Find the Fourier transform of f for t > 0 .
Problem 2.42. Find the Fourier transform of the function f (x) = x1 |x|α ,
where x1 is the first component of x and −n − 2 < α < −2 .
Hint: Use the fact that for −n < α < 0 ,
Γ( n+α
2 ) α+ n
F(|x|α )(ξ) = 2 2 |ξ|−(α+n)
Γ(− α2 )
1 ∂
and f (x) = |x|α+2 .
α + 2 ∂x1
Problem 2.43. Let α > 0 be given. Show that the Fourier transform of the
function
∞
1
Z
2
f (x) = tα−1 e−t e−t|x| dt
Γ(α) 0
is positive.
Problem 2.44. Let f ∈ L1 (R) . Show that the anti-derivative of f can be

written as the convolution of f and a function ϕ ∈ L1loc (R) .
Proof. Let ϕ be the characteristic function of the set (0, ∞) , or

1 if x > 0 ,
ϕ(x) =
0 if x ≤ 0 .
Then ϕ ∈ L1loc (R) , and

Z Z x
(ϕ ∗ f )(x) = ϕ(x − y)f (y)dy = f (y)dy
R −∞
which is the anti-derivative of f .
Problem 2.45. Let f be a continuous function with period 2π , and fˆ be

the Fourier transform of f . Show that
∞
√
fˆ(ξ) =
X
( 2πfn )τ−n δ
n=−∞
in the sense of distribution, where fn is the Fourier coefficient defined by

Z 2π
1
fn = f (x)e−inx dx .
2π 0
Problem 2.46. Using Definition 2.127, compute the Fourier transform of
the function/distribution
(
x if x ≥ 0 ,
R(x) =
0 otherwise ,
by completing the following:

1
(1) Let H be the Heaviside function. Show that Ĥ(ξ) = p.v. √ +Cδ(ξ)
2πiξ
1
for some constant C , where p.v. is defined as
ξ
D 1 E
Z
ϕ(ξ) Z − Z ∞ ϕ(ξ)
p.v. , ϕ = lim dξ = lim + dξ .
ξ →0+ R\[−,] ξ →0+ −∞ ξ
Note that the integral above always exists as long as ϕ ∈ S(R) .

1
(2) Let S(x) = H(x) − . Then S is an odd function, and show that
2
Ŝ(ξ) = −Ŝ(−ξ) .
(3) Use (2) to determine the constant C in (1).
(4) By the definition of Fourier transform, show that hR̂, ϕi = −ihĤ, ϕ0 i ,

and as a consequence
d
R̂(ξ) = i Ĥ(ξ) .
dξ
2.4. THE SOBOLEV SPACES H S (RN ), S ∈ R 93
2.4 The Sobolev Spaces H s (Rn ), s ∈ R

2.4.1 H s (Rn ) via the Fourier Transform
The Fourier transform allows us to generalize the Hilbert spaces H k (Rn ) for
k ∈ Z+ to H s (Rn ) for all s ∈ R, and hence study functions which possess
fractional derivatives (and anti-derivatives) which are square integrable.
p
Definition 2.142. For any s ∈ Rn , let hξi = 1 + |ξ|2 , and set
H s (Rn ) = {u ∈ S 0 (Rn ) | hξis û ∈ L2 (Rn )}

= {u ∈ S 0 (Rn ) | Λs u ∈ L2 (Rn )} ,
where Λs u = F ∗ (hξis û).
The operator Λs can be thought of as a “differential operator” of order

s, and according to Rellich’s theorem, Λ−s is a compact operator, yielding
the isomorphism
H s (Rn ) = Λ−s L2 (Rn ) .
Definition 2.143. The inner-product on H s (Rn ) is given by
(u, v)H s (Rn ) = (Λs u, Λs v)L2 (Rn ) ∀u, v ∈ H s (Rn ) .
and the norm on H s (Rn ) is
kuksH s (Rn ) = (u, u)H s (Rn ) ∀u ∈ H s (Rn ) .
The completeness of H s (Rn ) with respect to the k · kH s (Rn ) ) is induced

by the completeness of L2 (Rn ).
Theorem 2.144. For s ∈ R, (H s (Rn ), k · kH s (Rn ) ) is a Hilbert space.
Example 2.145 (H 1 (Rn )). The H 1 (Rn ) in Fourier representation is exactly

the same as the that given by Definition 2.73:
Z
2
kukH 1 (Rn ) = hξi2 kû(ξ)k2 dξ
R n
Z
= (1 + |ξ|2 )kû(ξ)k2 dξ
n
ZR
= (|u(x)|2 + |Du(x)|2 )dx ,
Rn
the last equality following from the Plancheral theorem.

1 1
Example 2.146 (H 2 (Rn )). The H 2 (Rn ) can be viewed as interpolating
between decay required for û ∈ L2 (Rn ) and û ∈ H 1 (Rn ):
Z p
1
n 2 n
H 2 (R ) = {u ∈ L (R ) | 1 + |ξ|2 |û(ξ)|2 dξ < ∞} .
Rn
Example 2.147 (H −1 (Rn )). The space H −1 (Rn ) can be heuristically de-
scribed as those distributions whose anti-derivative is in L2 (Rn ); in terms of
the Fourier representation, elements of H −1 (Rn ) possess a transforms that
can grow linearly at infinity:
|û(ξ)|2
Z
H −1 (Rn ) = {u ∈ S 0 (Rn ) | dξ < ∞} .
Rn 1 + |ξ|2
For T ∈ H −s (Rn ) and u ∈ H s (Rn ), the duality pairing is given by
hT, ui = (Λ−s T, Λs u)L2 (Rn ) ,
from which the following result follows.
Proposition 2.148. For all s ∈ R, [H s (Rn )]0 = H −s (Rn ) .
The ability to define fractional-order Sobolev spaces H s (Rn ) allows us to

refine the estimates of the trace of a function which we previously stated in
Theorem 2.93. That result, based on the Gauss-Green theorem, stated that
the trace operator was continuous from H 1 (Rn+ ) into L2 (Rn−1 ). In fact, the
1
trace operator is continuous from H 1 (Rn+ ) into H 2 (Rn−1 ).
To demonstrate the idea, we take n = 2. Given a continuous function
u : R2 → {x1 = 0}, we define the operator
T u = u(0, x2 ) .
The trace theorem asserts that we can extend T to a continuous linear map
1
from H 1 (R2 ) into H 2 (R) so that we only lose one-half of a derivative.
1
Theorem 2.149. T : H 1 (R2 ) → H 2 (R), and there is a constant C such
that
kT uk 1 ≤ CkukH 1 (R2 ) .
H 2 (R)
Before we proceed with the proof, we state a very useful result.

Lemma 2.150. Suppose that u ∈ S(R2 ) and define f (x2 ) = u(0, x2 ). Then
1
Z
ˆ
f (ξ2 ) = √ û(ξ1 , ξ2 )dξ1 .
2π Rξ1
Proof. fˆ(ξ2 ) = √1 √1 F ∗
R R
2π R û(ξ1 , ξ2 )dξ1 if and only if f (x2 ) = 2π R û(ξ1 , ξ2 )dξ1 ,
and
1 1
Z Z Z
√ F∗ û(ξ1 , ξ2 )dξ1 = û(ξ1 , ξ2 )dξ1 eix2 ξ2 dξ2 .
2π R 2π R R
On the other hand,
1
Z Z
∗
u(x1 , x2 ) = F [û(ξ1 , ξ2 )] = û(ξ1 , ξ2 )eix1 ξ1 +ix2 ξ2 dξ1 dξ2 ,
2π R R
so that
1
Z Z
u(0, x2 ) = F ∗ [û(ξ1 , ξ2 )] = û(ξ1 , ξ2 )eix2 ξ2 dξ1 dξ2 .
2π R R
Proof of Theorem 2.149. Suppose that u ∈ S(R2 ) and set f (x2 ) = u(0, x1 ).
According to Lemma 2.150,
1 1
Z Z
ˆ
f (ξ2 ) = √ û(ξ1 , ξ2 )dξ1 = √ û(ξ1 , ξ2 )hξi hξi−1 dξ1
2π Rξ1 2π Rξ1
Z 1 Z 1
1 2 2
2
−2
2
≤√ |û(ξ1 , ξ2 )| hξi dξ1 hξi dξ1 ,
2π R R
and hence Z Z
|f (ξ2 )| ≤ C 2 2
|û(ξ1 , ξ2 )| hξi dξ1 2
hξi−2 dξ1 .
R R
The key to this trace estimate is the explicit evaluation of the integral
−2
R
R hξi dξ1 :
+∞
tan−1 √ ξ1 2

1
Z
1+ξ2 1
≤ π(1 + ξ22 )− 2 .

dξ1 = (2.47)
1 + ξ12 + ξ22
p
R 1 + ξ22

−∞
2 21 ˆ 2 2 2
R R
It follows that R (1 + ξ2 ) |f (ξ2 )| dξ2 ≤C R |û(ξ1 , ξ2 )| hξi dξ1 , so that inte-
gration of this inequality over the set {ξ2 ∈ R} yields the result. Using the
density of S(R2 ) in H 1 (R2 ) completes the proof.
The proof of the trace theorem in higher dimensions and for general
H s (Rn ) spaces, s > 12 , replacing H 1 (Rn ) proceeds in a very similar fashion;
the only difference is that the integral R hξi−2 dξ1 is replaced by Rn−1 hξi−2s dξ1 ·
R R
· · dξn−1 , and instead of obtaining an explicit anti-derivative of this integral,

an upper bound is instead found. The result is the following general trace
theorem.
1
Theorem 2.151 (The trace theorem for H s (Rn )). For s > 2, the trace
s− 12
operator T : H s (Rn ) →H (Rn−1 ) is continuous.
We can extend this result to open, bounded, C ∞ domains Ω ⊂ Rn .
Definition 2.152. Let ∂Ω denote a closed C ∞ manifold, and let {ωl }K

l=1
denote an open covering of ∂Ω, such that for each l ∈ {1, 2, ..., K}, there
exist C ∞ -class charts ϑl which satisfy
ϑl : B(0, rl ) ⊂ Rn−1 → ωl is a C ∞ diffeomorphism .
Next, for each 1 ≤ l ≤ K, let 0 ≤ ϕl ∈ C0∞ (Ul ) denote a partition of unity

so that L
P
l=1 ϕl (x) = 1 for all x ∈ ∂Ω. For all real s ≥ 0, we define
H s (∂Ω) = {u ∈ L2 (∂Ω) : kukH s (∂Ω) < ∞} ,
where for all u ∈ H s (∂Ω),

K
X
kuk2H s (∂Ω) = k(ϕl u) ◦ ϑl k2H s (Rn−1 ) .
l=1
The space (H s (∂Ω), k · kH s (∂Ω) ) is a Hilbert space by virtue of the com-

pleteness of H s (Rn−1 ); furthermore, any system of charts for ∂Ω with sub-
ordinate partition of unity will produce an equivalent norm.
1
Theorem 2.153 (The trace map on Ω). For s > 2, the trace operator
T : Ω → ∂Ω is continuous.
Proof. Let {Ul }K

l=1 denote an n-dimensional open cover of ∂Ω such that
Ul ∩ ∂Ω = ωl . Define charts θl : Vl → Ul , as in (2.17) but with each
chart being a C ∞ map, such that ϑl is equal to the restriction of θl to the
(n − 1)-dimensional ball B(0, rl ) ⊂ Rn−1 ). Also, choose a partition of unity
0 ≤ ζl ∈ C0∞ (Ul ) subordinate to the covering Ul such that ϕl = ζl |ωl .
Then by Theorem 2.151, for s > 12 ,
K
X K
X
kuk2 1 = k(ϕl u) ◦ ϑl k2 1 ≤C k(ϕl u) ◦ ϑl k2H s (Rn ) ≤ Ckuk2H s (Ω) .
H s− 2 (∂Ω) H s− 2 (Rn−1 )
l=1 l=1
One may then ask if the trace operator T is onto; namely, given f ∈
s− 12
H (Rn−1 ) for s > 12 , does there exist a u ∈ H s (Rn ) such that f = T u?
By essentially reversing the order of the proof of Theorem 2.149, it is possible
to answer this question in the affirmative. We first consider the case that
n = 2 and s = 1.
1
Theorem 2.154. T : H 1 (R2 ) → H 2 (R) is a surjection.
Proof. With ξ = (ξ1 , ξ2 ), we define (one of many possible choices) the func-
tion u on R2 via its Fourier representation:
hξ1 i
û(ξ1 , ξ2 ) = K fˆ(ξ1 ) 2 ,
hξi
for a constant K 6= 0 to be determined shortly. To verify that kukH 1 (R1 ) ≤

kf k 1 , note that
H 2 (R)
∞ ∞ ∞ ∞
hξ1 i2
Z Z Z Z
2 2
|û(ξ1 , ξ2 )| hξi dξ1 dξ2 = K |fˆ(ξ1 )|2
dξ1 dξ2
−∞ −∞ −∞ −∞ hξi2
Z ∞ Z ∞
1
=K |fˆ(ξ1 )|2 (1 + ξ12 ) 2 2 dξ2 dξ1
−∞ −∞ 1 + ξ 1 + ξ2
≤ Ckf k2 1 ,
H 2 (R)
where we have used the estimate (2.47) for the inequality above.
It remains to prove that u(x1 , 0) = f (x1 ), but by Lemma 2.150, it suffices
that Z ∞ √
û(ξ1 , ξ2 )dξ2 = 2π fˆ(ξ1 ) .
−∞
Integrating û, we find that

Z ∞ ∞
1
q Z
ˆ 2
û(ξ1 , ξ2 )dξ2 = K f (ξ1 ) 1 + ξ1 dξ2 ≤ Kπ fˆ(ξ1 )
−∞ −∞ 1 + ξ12 + ξ22
√
so setting K = 2π/π completes the proof.
A similar construction yields the general result.

1
Theorem 2.155. For s > 12 , T : H s (Rn ) → H s− 2 (Rn−1 ) is a surjection.
By using the system of charts employed for the proof of Theorem 2.153,
we also have the surjectivity of the trace map on bounded domains.
1
Theorem 2.156. For s > 12 , T : H s (Ω) → H s− 2 (∂Ω) is a surjection.
The Fourier representation provides a very easy proof of a simple version

of the Sobolev embedding theorem.
Theorem 2.157. For s > n/2, if u ∈ H s (Rn ), then u is continuous and
max |u(x)| ≤ CkukH s (Rn ) .
Proof. By Theorem 2.118, u = F ∗ û; thus according to Hölder’s inequality

and the Riemann-Lebesgue lemma (Theorem 2.134), it suffices to show that
kûkL1 (Rn ) ≤ CkukH s (Rn ) .
But this follows from the Cauchy-Schwarz inequality since

Z Z
|û(ξ)|dξ = |û(ξ)|hξis hξi−s dξ
Rn Rn
Z 1 Z 1
2 2
2 2s −2s
≤ |û(ξ)| hξi dξ hξi dξ
Rn Rn
≤ CkukH s (Rn ) ,
the latter inequality holding whenever s > n/2.
Hölder’s inequality can be used to prove the following
Theorem 2.158 (Interpolation inequality). Let 0 < r < t < ∞, and s =

αr + (1 − α)t for some α ∈ (0, 1). Then
kukH s (Rn ) ≤ CkukαH r (Rn ) kuk1−α

H t (Rn ) . (2.48)
Example 2.159 (Euler equation on T2 ). On some time interval [0, T ] sup-

pose that u(x, t), x ∈ T2 , t ∈ [0, T ], is a smooth solution of the Euler equa-
tions:
∂t u + (u · D)u + Dp = 0 in T2 × (0, T ] ,
div u = 0 in T2 × (0, T ] ,
with smooth initial condition u|t=0 = u0 . Written in components, u =

(u1 , u2 ) satisfies uit + ui ,j j j + p,i = 0 for i = 1, 2, where we are using the
Einstein summation convention for summing repeated indices from 1 to 2
and where ui ,j = ∂ui /∂xj and p,i = ∂p/∂xi .
Computing the L2 (T2 ) inner-product of the Euler equations with u yields
the equality
1d
Z Z Z
2 i j i
|u(x, t)| dx + u ,j u u dx + p,i ui dx = 0 .
2 dt T2 T2 T2
| {z } | {z }
I1 I2
Notice that
1 1
Z Z
2 j
I1 = (|u| ),j u dx = |u|2 div udx = 0 ,
2 T2 2 T2
the second equality arising from integration by parts with respect to ∂/∂xj .
Integration by parts in the integral I2 shows that I2 = 0 as well, from which
d 2
the conservation law dt ku(·, t)kL2 (T2 ) follows.
To estimate the rate of change of higher-order Sobolev norms of u relies
on the use of the Sobolev embedding theorem. In particular, we claim that
on a short enough time interval [0, T ], we have the inequality
d
ku(·, t)k2H 3 (T2 ) ≤ Cku(·, t)k3H 3 (T2 ) (2.49)
dt
from which it follows that ku(·, t)k2H 3 (T2 ) ≤ M for some constant M < ∞.
To prove (2.49), we compute the H 3 (T2 ) inner-product of the Euler equa-
tions with u:
1d X Z X Z
ku(·, t)k2H 3 (T2 ) + Dα ui ,j uj Dα ui dx+ Dα p,i Dα ui dx = 0 .
2 dt T2 T2
|α|≤3 |α|≤3
The third integral vanishes by integration by parts and the fact that Dα div u =
0; thus, we focus on the nonlinearity, and in particular, on the highest-order
derivatives |α| = 3, and use D3 to denote all third-order partial derivatives,
as well as the notation l.o.t. for lower-order terms. We see that
Z Z
3 i j 3 i
D (u ,j u )D u dx = D3 ui ,j uj D3 ui dx
T2 T2
| {z }
K1
Z Z
i 3 j 3 i
+ u ,j D u D u dx + l. o. t. dx .
T2 T2
| {z }
K2
l. o. t. dx ≤ Ckuk3H 3 (T2 ) , so
R
By definition of being lower-order terms, T2
it remains to estimate the integrals K1 and K2 . But the integral K1 vanishes
by the same argument that proved I1 = 0. On the other hand, the integral
K2 is estimated by Hölder’s inequality:
|K2 | ≤ kui ,j kL∞ (T2 ) kD3 uj kH 3 (T2 ) kD3 ui kH 3 (T2 ) .
Thanks to the Sobolev embedding theorem, for s = 2 (s needs only to be

greater than 1),
kui ,j kL∞ (T2 ) ≤ Ckui ,j kH 2 (T2 ) ≤ kukH 3 (T2 ) ,
from which it follows that K2 ≤ Ckuk3H 3 (T2 ) , and this proves the claim.
Note well, that it is the Sobolev embedding theorem that requires the use
of the space H 3 (T2 ) for this analysis; for example, it would not have been
possible to establish the inequality (2.49) with the H 2 (T2 ) norm replacing
the H 3 (T2 ) norm.
2.4.2 Fractional-order Sobolev spaces via difference quotient

norms
The case that s > 0
Lemma 2.160. For 0 < s < 1, u ∈ H s (Rn ) is equivalent to
|u(x) − u(y)|2
ZZ
2 n
u ∈ L (R ) , dxdy < ∞ .
Rn ×Rn |x − y|n+2s
Proof. The Fourier transform shows that for h ∈ Rn ,
h·ξ
Z Z Z
2 ih·ξ 2 2
|u(x + h) − u(x)| dx = |e − 1| |û(ξ)| dξ = sin2 |û(ξ)|2 dξ .
Rn Rn Rn 2
It follows that
|u(x) − u(y)|2 sin2 h·ξ

ZZ ZZ
dxdy = 2
|û(ξ)|2 dξdh
Rn ×Rn |x − y|n+2s n
R ×R n |h| n+2s
Z h Z sin2 h·ξ i
2 2
= |û(ξ)| n+2s
dh dξ
Rn Rn |h|
Z h Z sin2 (z · ξ ) i
−2s 2s 2 |ξ|
(letting h = 2|ξ|
−1
z) = 2 |ξ| |û(ξ)| n+2s
dz dξ .
R n Rn |z|
As the integral inside of the square brackets is rotationally invariant, it is

independent of the direction of ξ/|ξ|; as such we set ξ/|ξ| = e1 and let
z1 = z · e1 denote the first component of the vector z. It follows that
|u(x) − u(y)|2
ZZ Z
dxdy = C |ξ|2s |û(ξ)|2 dξ ,
Rn ×Rn |x − y|n+2s Rn
sin2 z1
Z
where C = dz < ∞ .
Rn |z|n+2s
Corollary 2.161. For 0 < s < 1,

h ZZ |u(x) − u(y)|2 i1
2
kukH s (Rn ) = kukL2 (Rn ) + dxdy
is an equivalent norm on H s (Rn ).
For real s ≥ 0, u ∈ H s (Rn ) if and only if Dα u ∈ L2 (Rn ) for all |α| ≤ [s]
(where [s] denotes the greatest integer that is not bigger than s), and
|Dα u(x) − Dα u(y)|2

ZZ
dxdy < ∞
Rn ×Rn |x − y|n+2(s−[s])
for all |α| = [s]. Moreover, an equivalent norm on H s (Rn ) is given by
|Dα u(x) − Dα u(y)|2 i1

h X ZZ
α 2
kukH s (Rn ) = kD uk2L2 (Ω(Rn ) + n+2(s−[s])
dxdy .
Rn ×Rn |x − y|
|α|≤[s]
If u ∈ H k (Rn ), k ∈ N, and ϕ ∈ S(Rn ), application of the product

rule shows that ϕu ∈ H k (Rn ). When s 6∈ N, however, the product rule
is not directly applicable and we must rely on other means to show that
ϕu ∈ H s (Rn ).
Lemma 2.162. Suppose that u ∈ H s (Rn ) for some s ≥ 0 and ϕ ∈ S(Rn ).

Then ϕu ∈ H s (Rn ).
Proof. We first conisder the case that 0 ≤ s < 1.

By Corollary 2.160, since ϕu is clearly an L2 (Rn )-function, it suffices to
show that
|(ϕu)(x) − (ϕu)(y)|2
ZZ
dxdy < ∞ .
Since |(ϕu)(x) − (ϕu)(y)| ≤ |ϕ(x) − ϕ(y)||u(x)| + |u(x) − u(y)||ϕ(y)|,

|(ϕu)(x) − (ϕu)(y)|2
ZZ
dxdy
|ϕ(x) − ϕ(y)|2 |u(x)|2 + |u(x) − u(y)|2 |ϕ(y)|2
ZZ
≤2 dxdy
|ϕ(x) − ϕ(y)|2 |u(x)|2 |u(x) − u(y)|2
ZZ ZZ
≤2 dxdy + 2kϕk ∞
L (R )n dxdy .
Rn ×Rn |x − y|n+2s Rn ×Rn |x − y|
n+2s
| {z } | {z }
I1 I2
Since u ∈ H s (Rn ), I2 < ∞. On the other hand,

hZ Z Z Z i |ϕ(x) − ϕ(y)|2 |u(x)|2
I1 = + dxdy .
Rn |x−y|≤1 Rn |x−y|≥1 |x − y|n+2s
For the integral over |x − y| ≤ 1, since ϕ ∈ S(Rn ), |ϕ(x) − ϕ(y)| ≤ C|x − y|

for some constant C. Therefore,
|ϕ(x) − ϕ(y)|2 |u(x)|2
Z Z Z Z
dxdy ≤ C |x − y|2−n−2s |u(x)|2 dxdy
Rn |x−y|≤1 |x − y|n+2s Rn |x−y|≤1
Z Z
≤C |z|2−n−2s dz |u(x)|2 dx < ∞ if s < 1 .
|z|≤1 Rn
For the remaining integral,

|ϕ(x) − ϕ(y)|2 |u(x)|2
Z Z Z Z
2
dxdy ≤ 4kϕkL∞ (Rn ) |x − y|−n−2s |u(x)|2 dxdy
Rn |x−y|≥1 |x − y|n+2s Rn |x−y|≤1
Z Z n
2 −n−2s
≤ 4kϕkL∞ (Rn ) |z| dz |u(x)|2 < ∞ if s > 0 .
|z|≥1 R
The general case of s ≥ 0 can be proved in a similar fashion, and we
leave the details to the reader.
The case that s < 0
For s < 0, we define the space H s (Rn ) to be the dual space of H −s (Rn ) with
the corresponding dual space norm (or operator norm) defined by
hu, vi
kukH s (Rn ) = sup = sup hu, vi . (2.50)
v∈H −s (Rn ) kvkH −s (Rn ) kvkH −s (Rn ) =1
The norm defined in (2.50) is equivalent to

hZ i1
2
kukH s (Rn ) = (1 + |ξ|2s )|û(ξ)|2 dξ .
Rn
2.5. FRACTIONAL-ORDER SOBOLEV SPACES ON DOMAINS WITH BOUNDARY103
2.5 Fractional-order Sobolev spaces on domains

with boundary
2.5.1 The space H s (Rn+ )
Let Rn+ = Rn−1 × R+ denote the upper half space of Rn .
The case s = k ∈ N
The space H k (Rn+ ) is the collection of all L2 (Rn+ )-functions so that the α-th
weak derivatives belong to L2 (Rn+ ) for all |α| ≤ k, that is,
n o
H k (Rn+ ) = u ∈ L2 (Rn+ ) Dα u ∈ L2 (Rn+ ) ∀ |α| ≤ k

with norm
X
kuk2H k (Rn ) = kDα uk2L2 (Rn ) . (2.51)
+ +
|α|≤k
Note that we are not able to directly use the Fourier transform to define the
H k (Rn+ ).
Definition 2.163 (Extension operator E). Fix N ∈ N. Let (a1 , · · · , aN )

solve
N
X
(−j)` aj = 1 , ` = 0, · · · , N − 1 .
j=1
We denote by E : C(Rn+ ) → C(Rn ) the function
u(x) xn ≥ 0 ,

 if
N
(Eu)(x) = (2.52)
aj u(x0 , −jxn )
P
 if xn < 0 .
j=1
Note that the the coefficients aj solve a linear system of N equations for
N unknowns which is always solvable since the determinant never vanishes.
Theorem 2.164 (Sobolev extension theorem). The operator E has a con-

tinuous extension to an operator E : H k (Rn+ ) → H k (Rn ), k ≤ N − 1, with
N defined in (2.52).
Proof. We must show that all derivatives of u of order not bigger than N − 1
are continuous at xn = 0. We compute Dx` n Eu:
Dx` n u(x) if xn > 0 ,



Dx` n (Eu)(x) = N
(−j)` aj (Dx` n u)(x0 , −jxn )
P
 if xn < 0 .
j=1
By the definition of aj , lim Dx` 1 (Eu)(x) = lim Dx` 1 (Eu)(x). So Eu ∈

x1 →0+ x1 →0−
H k (Rn ). Finally, the continuity of E is concluded by the following inequal-
ity:
kEukH k (Rn ) ≤ CkukH k (Rn+ ) .
Lemma 2.165. For k ∈ N, each u ∈ H k (Rn+ ) is the restriction of some

w ∈ H k (Rn ) to Rn+ , that is, u = w|Rn+ .
Proof. We define the restriction map % : H k (Rn ) → H k (Rn+ ). By Theorem

2.164, the restriction map is onto, since %E = Id on H k (Rn+ ).
The case s 6∈ N
Next, suppose that N − 2 < s < N − 1 for some N ∈ N given in (2.52), and
let E continue to denote the Sobolev extension operator.
We define the space H s (Rn+ ) as the restriction of H s (Rn ) to Rn+ with
norm
kukH s (Rn+ ) ≡ kEukH s (Rn ) . (2.53)
When s = k ∈ N, it may not be immediately clear that the H s (Rn+ )-norm

defined by (2.53) is equivalent to the H k (Rn+ )-norm defined by (2.51). Let
k · k1 be the norm defined by (2.51) and k · k2 be the norm defined by (2.53).
It is clear that kuk1 ≤ kuk2 , and by the continuity of E, kuk2 ≤ Ckuk1 ;
therefore, k · k1 and k · k2 are equivalent if s ∈ N.
2.5.2 The Sobolev space H s (Ω)
We can now define the Sobolev spaces H s (Ω) for any open and bounded
domain Ω ⊂ Rn with smooth boundary ∂Ω.
2.5. FRACTIONAL-ORDER SOBOLEV SPACES ON DOMAINS WITH BOUNDARY105
Definition 2.166 (Smoothness of the boundary). We say ∂Ω is C k if for

each point x0 ∈ ∂Ω there exist r > 0 and a C k -function γ : Rn−1 → R such
that - upon relabeling and reorienting the coordinates axes if necessary - we
have
Ω ∩ B(x0 , r) = {x ∈ B(x0 , r) | xn > γ(x1 , · · · , xn−1 )} .
∂Ω is C ∞ if ∂Ω is C k for all k ∈ N, and Ω is said to have smooth boundary

if ∂Ω is C ∞ .
Definition 2.167 (Partition of unity). Let X be a topological space. A

partition of unity is a collection of continuous functions {χj : X → [0, 1]}
P
such that j χi (x) = 1 for all x ∈ X. A partition of unity is locally finite
if each x in X is contained in an open set on which only a finite number of
χj are non-zero. A partition of unity is subordinate to an open cover {Uj }
of X if each χi is zero on the complement of Uj .
For a domain Ω with smooth boundary, we may assume that there exist
x1 , · · · xN ∈ ∂Ω, r1 , · · · rN > 0, γj ∈ C ∞ such that, upon relabeling and
reorienting the coordinates axes if necessary,
Ω ∩ Uj = {x ∈ Uj | xn > γj (x1 , · · · , xn−1 )} where Uj = B(xj , rj ) ,

SN
and Ω ⊆ j=0 Uj , and {χj }N
j=0 is a partition of unity subordinate to the
open cover {Uj } such that χj ∈ Cc∞ (Uj ), and the function ψj defined by
ψj (x) = (x1 , · · · , xn−1 , γj (x1 , · · · , xn−1 ) + xn ) .
is a diffeomorphism between a small neighborhood Vj of Rn and supp(χj ).

Let v0 = χ0 u and vj = (χj u) ◦ ψj . Then v0 can be considered as a
function defined on Rn , and vj can be considered as a function defined on
Rn+ . We then have the following definition.
Definition 2.168. The space H s (Ω) for s > 0 is the collection of all mea-
surable functions u such that χ0 u ∈ H s (Rn ) and (χj u) ◦ ψj ∈ H s (Rn+ ). The
H s (Ω)-norm is defined by
h N
X i1/2
kukH s (Ω) = kχ0 uk2H s (Rn ) + k(χj u) ◦ ψj k2H s (Rn ) .
+
j=1
Theorem 2.169 (Extension). Let Ω be a bounded, smooth domain. For

any open set V such that Ω ⊂⊂ U , there exists a bounded linear operator
E : H s (Ω) → H s (Rn ) such that
(i) Eu = u a.e. in Ω ,
(ii) Eu has support within V ,
(iii) kEukH s (Rn ) ≤ CkukH s (Ω) , where the constant C depends only on s, Ω
and V .
Proof. Define
N
X √ h √ i
Eu = χ0 u + χj E[( χj u) ◦ ψj ] ◦ ψj−1 ,
j=1
where E : H k (Rn+ ) → H k (Rn ) is the continuous extension defined by (2.52)

for some k ≥ s. One more constraint, supp(χj ) ⊆ V , must be imposed on
χj for all j because of (ii), while this constraint is easily satisfied if we let
rj ≤ dist(Ω, ∂U ) for all j.
2.6 The Sobolev Spaces H s (Tn ), s ∈ R

2.6.1 The Fourier Series: Revisited
Definition 2.170. For u ∈ L1 (Tn ), define

Z
Fu(k) = ûk = (2π)−n e−ik·x u(x)dx ,
Tn
and
X
F ∗ û(x) = ûk eik·x .
k∈Zn
Note that F : L1 (Tn ) → l∞ (Zn ). If u is sufficiently smooth, then inte-
gration by parts yields
F(Dα u) = −(−i)|α| k α ûk , k α = k1α1 · · · knαn .
Example 2.171. Suppose that u ∈ C 1 (Tn ). Then for j ∈ {1, ..., n},

∂u ∂u −ik·x
Z
−n
F (k) = (2π) e dx
∂xj Tn ∂xj
Z
= −(2π)−n u(x) (−ikj ) e−ik·x dx
Tn
= ikj ûk .
2.6. THE SOBOLEV SPACES H S (TN ), S ∈ R 107
Note that Tn is a closed manifold without boundary; alternatively, one may

identify Tn with the [0, 1]n with periodic boundary conditions, i.e., with op-
posite faces identified.
Definition 2.172. Let s = S(Zn ) denote the space of rapidly decreasing

functions û on Zn such that for each N ∈ N,
pN (u) = sup hkiN |ûk | < ∞ ,

k∈Zn
p
where hki = 1 + |k|2 .
Then
F : C ∞ (Tn ) → s , F ∗ : s → C ∞ (Tn ) ,
and F ∗ F = Id on C ∞ (Tn ) and FF ∗ = Id on s. These properties smoothly

extend to the Hilbert space setting:
F : L2 (Tn ) → l2 F ∗ : l2 → L2 (Tn )
F ∗ F = Id on L2 (Tn ) FF ∗ = Id on l2 .
Definition 2.173. The inner-products on L2 (Tn ) and l2 are

Z
n
(u, v)L2 (Tn ) = (2π)− 2 u(x)v(x)dx
Tn
and
X
(û, v̂)l2 = ûk v̂k ,
k∈Zn
respectively.
Parseval’s identity shows that kukL2 (Tn ) = kûkl2 .
Definition 2.174. We set
D0 (Tn ) = [C ∞ (Tn )]0 and s0 = [s]0 .
The space D0 (Tn ) is termed the space of periodic distributions.
In the same manner that we extended the Fourier transform from S(Rn )
to S 0 (Rn ) by duality, we may produce a similar extension to the periodic
distributions:
F : D0 (Tn ) → s0 F ∗ : s0 → D0 (Tn )
F ∗ F = Id on D0 (Tn ) FF ∗ = Id on s0 .
Definition 2.175 (Sobolev spaces H s (Tn )). For all s ∈ R, the Hilbert
spaces H s (Tn ) are defined as follows:
H s (Tn ) = {u ∈ D0 (Tn ) | kukH s (Tn ) < ∞} ,
where the norm on H s (Tn ) is defined as

X
kuk2H s (Tn ) = |ûk |2 hki2s .
k∈Zn
The space (H s (Tn ), k · kH s (Tn ) ) is a Hilbert space, and we have that
H −s (Tn ) = [H s (Tn )]0 .
For any s ∈ R, we define the operator Λs as follows: for u ∈ D0 (Tn ),

X
Λs u(x) = |ûk |2 hkis eik·x .
k∈Zn
It follows that
H s (Tn ) = Λ−s L2 (Tn ) ,
and for r, s ∈ R,
Λs : H r (Tn ) → H r−s (Tn ) is an isomorphism .
Notice then that for any > 0,
Λ− : H s (Tn ) → H s (Tn ) is a compact operator ,
as it is an operator-norm limit of finite-rank operators. (In particular, the

eigenvalues of Λ− tend to zero in this limit.) Hence, the inclusion map
H s+ (Tn ) ,→ H s (Tn ) is compact, and we have the following
Theorem 2.176 (Rellich’s theorem on Tn ). Suppose that a sequence uj

satisfies for s ∈ R and > 0,
sup kuj kH s+ (Tn ) ≤ M < ∞ for a constant M 6= M (j) .
Then there exists a subsequence ujk → u in H s (Tn ).

2.6.2 The Poisson Integral Formula and the Laplace operator
For f : S1 → R, denote by PI(f )(r, θ) the harmonic function on the unit

disk D = {x ∈ R2 : |x| < 1} with trace f :
∆ PI(f ) = 0 in D
PI(f ) = f on ∂D = S1 .
PI(f ) has an explicit representation via the Fourier series
fˆk r|k| eikθ r < 1, 0 ≤ θ < 2π ,

X
PI(f )(r, θ) = (2.54)
k∈Z
as well as the integral representation
1 − r2 f (φ)
Z
PI(f )(r, θ) = dφ r < 1, 0 ≤ θ < 2π .
2π S1 r2 − 2r cos(θ − φ) + 1
(2.55)
The dominated convergence theorem shows that if f ∈ C 0 (S1 ), then PI(f ) ∈

C ∞ (D) ∩ C 0 (D).
1
Theorem 2.177. PI extends to a continuous map from H k− 2 (S1 ) to H k (D)
for all k ∈ Z+ .
Proof. Define u = PI(f ).

1
Step 1. The case that k = 0. Assume that f ∈ H − 2 (Γ) so that
|fˆk |2 hki−1 ≤ M0 < ∞ .

X
k∈Z
Since the functions {r|k| eikθ : k ∈ Z} are orthogonal with respect to the
L2 (D) inner-product,
2
Z 2π Z 1 X
kuk2L2 (D) = fˆk r e|k| ikθ
r dr dθ

0 0
k∈Z

Z 1
|fˆk |2 |fˆk |2 (1 + |k|)−1 ≤ πkf k2
X X
≤ 2π r2|k|+1 dr = π 1 ,
0 H 2 (S1 )
k∈Z k∈Z
where we have used the monotone convergence theorem for the first inequal-
ity.
1
Step 2. The case that k = 1. Next, suppose that f ∈ H 2 (Γ) so that
|fˆk |2 hki1 ≤ M1 < ∞ .

X
k∈Z
Since we have shown that u ∈ L2 (D), we must now prove that uθ = ∂θ u

and ur = ∂r u are both in L2 (D). Notice that by definition of the Fourier
transform and (2.54),
∂
PI(f ) = PI(fθ ) . (2.56)
∂θ
1 1
By definition, ∂θ : H 2 (S1 ) → H − 2 (S1 ) continuously, so that for some con-
stant C,
kfθ k 1 ≤ Ckf k 1 .
H − 2 (S1 ) H 2 (S1 )
It follows from the analysis of Step 1 and (2.56) that (with u = PI(f )),
kuθ kL2 (D) ≤ Ckf k 1 .

H 2 (S1 )
Next, using the identity (2.54) notice that |rur | = |uθ |. It follows that
krur kL2 (D) ≤ Ckf k 1 . (2.57)

H 2 (S1 )
By the interior regularity of −∆ proven in Theorem 2.178, ur (r, θ) is smooth

on {r < 1}; hence the bound (2.57) implies that, in fact,
kur kL2 (D) ≤ Ckf k 1 ,

H 2 (S1 )
and hence
kukH 1 (D) ≤ Ckf k 1 ,
H 2 (S1 )
1
Step 3. The case that k ≥ 2. Since f ∈ H k− 2 (S1 ), it follows that
k∂θk f k 1 ≤ Ckf k 1
H − 2 (S1 ) H k− 2 (S1 )
and by repeated application of (2.56), we find that
kukH k (D) ≤ Ckf k 1 .

H k− 2 (S1 )

The Hölder spaces on D are defines as follows: if u : D → R is bounded

and continuous, we write
kukC(D) := sup |u(x)| .

x∈D
For 0 < α ≤ 1, the αth -Hölder seminorm of u is

|u(x) − u(y)|

[u]C 0,α (D) := sup
x,y∈D,x6=y |x − y|α
and the αth -Hölder norm of u is
kukC 0,α (D) = kukC(D) + [u]C 0,α (D) .
According to Morrey’s inequality, if f ∈ H 3/2 (S 1 ) then for 0 < α < 1, f ∈

C 0,α (S1 ). Next, we use the result of Problem 2.49, together with Morrey’s
inequality (once again) and Theorem 2.90 to prove that u ∈ C 0,α (D). Let
us explain this.
We first prove the following:
f ∈ H 3/2 (S1 ) implies that f ∈ H 1/2+α (S1 ) for α ∈ (0, 1) which implies that f ∈ C 0,α (S1 ) ,
the last assertion meaning that |f (x + y) − f (x)| ≤ C|y|α .

We start with the identity

X
|f (x + y) − f (y)| = fˆk eikx (eiky − 1)

k∈Z

fˆk e (e − 1)
X ikx iky

=
k6=0
 1  1
2 2
ˆ
X X
2 1+2α   iky 2 −1−2α 
≤  |fk | hki |e − 1| hki
k6=0 k6=0
 1
2
X
iky 2 −1−2α 
= kf kH 1/2+α  |e − 1| hki .
k6=0
1
We consider |y| ≤ 2 and break the sum into two parts:
X X X
|eiky − 1|2 hki−1−2α = |eiky − 1|2 hki−1−2α + |eiky − 1|2 hki−1−2α .
k6=0 1 1
|k|≤ |y| |k|≥ |y|
For the second sum, we use that |eiky − 1|2 ≤ 4 and employ the integral test
R∞
to see that 1/|y| r−1−2α dr ≤ C|y|2α . For the first sum, we Taylor expand
about y = 0: eiky − 1 = iky + O(y 2 ). Once again, we employ the integral
test:
X Z 1/|y|
|eiky − 1|2 hki−1−2α ≤ |y|2 + |y|2 r2 r−1−2α dr ≤ C(|y|2 + |y|2α ) .
1 1
|k|≤ |y|
Since |y| ≤ 1/2, we see that

X
|eiky − 1|2 hki−1−2α ≤ C|y|α
k6=0
as α < 1.
Next, according to Theorem 2.177, if f ∈ H 3/2 (S1 ), then u solves −∆u =
0 in D with u = f on ∂D, and kukH 2 (D) ≤ Ckf kH 3/2 (S1 ) . By Theorem 2.90,
√
kDukLq (D) ≤ C qkukH 2 (D) ∀q ∈ [1, ∞).
Hence, by Morrey’s inequality, we see that u ∈ C 0,1−2/q (D), and thus in

C 0,α (D) for α ∈ (0, 1).
2.6.3 Exercises
Problem 2.47. Let D := B(0, 1) ⊂ R2 and let u satisfy the Neumann

problem
∆u = 0 in D , (2.58a)
∂u
= g in ∂D := S1 . (2.58b)
∂r
If u = P I(f ) :=
P ˆ |k| eikθ , show that for f ∈ H 3/2 (S1 ),
k∈Z fk r
g = N f, (2.59)
which is the same as

ĝk = |k|fˆk .
P ˆ ikθ
N denotes the Dirichlet to Neumann map given by N f (θ) = k∈Z fk |k|e
∂
or N f = −i ∂θ Hf = −iH ∂f∂θ , where H is the Hilbert transform, defined by
Hu(θ) = k∈Z (sgn k)ĝk eikθ .
P
2.7. REGULARITY OF THE LAPLACIAN ON Ω 113
−1 ikθ . Show
P
Problem 2.48. Define the function K(θ) = k6=0 |k| e that
K ∈ L2 (S1 ) ⊂ L1 (S1 ). Next, show that if g ∈ L2 (S1 ) and S1 g(θ)dθ =
R
0, a
solution to (2.59) is given by f (θ) = (2π)−1 S1 K(θ − φ)g(φ)dφ.
R
Problem 2.49. Consider the solution to the Neumann problem (2.58a) and
(2.58b). Show that g ∈ H 1/2 (S1 ) implies that u ∈ H 2 (D) and that

kuk2H 2 (D) ≤ C kgk2H 1/2 (S1 ) + kuk2L2 (D) .
2.7 Regularity of the Laplacian on Ω
We have studied the regularity properties of the Laplace operator on D =

B(0, 1) ⊂ R2 using the Poisson integral formula. These properties continue
to hold on more general open, bounded, C ∞ subsets Ω of Rn .
We revisit the Dirichlet problem
∆u = 0 in Ω , (2.60a)
u = f on ∂Ω . (2.60b)
1
Theorem 2.178. For k ∈ N, given f ∈ H k− 2 (∂Ω), there exists a unique
solution u ∈ H k (Ω) to (2.60) satisfying
kukH k (Ω) ≤ Ckf k 1 , C = C(Ω) .

H k− 2 (∂Ω)
Proof. Step 1. k = 1. We begin by converting (2.60) to a problem with ho-

mogeneous boundary conditions. Using the surjectivity of the trace operator
provided by Theorem 2.156, there exists F ∈ H 1 (Ω) such that T (F ) = f
on ∂Ω, and kF kH 1 (Ω) ≤ Ckf k 1 . Let U = u − F ; then U ∈ H 1 (Ω)
H 2 (∂Ω)
and by linearity of the trace operator, T (U ) = 0 on ∂Ω. It follows from
Theorem 2.95 that U ∈ H01 (Ω) and satisfies −∆U = ∆F in H01 (Ω); that is
h−∆U, vi = h∆F, vi for all v ∈ H01 (Ω).
According to Remark 2.106, −∆ : H01 (Ω) → H −1 (Ω) is an isomorphism,
so that ∆F ∈ H −1 (Ω); therefore, by Theorem 2.105, there exists a unique
weak solution U ∈ H01 (Ω), satisfying
Z
DU · Dv dx = h∆F, vi ∀v ∈ H01 (Ω) ,
Ω
with
kU kH 1 (Ω) ≤ Ck∆F kH −1 (Ω) , (2.61)
and hence
u = U + F ∈ H 1 (Ω) and kukH 1 (Ω) ≤ kf k 1 .

H 2 (∂Ω)
Step 2. k = 2. Next, suppose that f ∈ H 1.5 (∂Ω). Again employing

Theorem 2.156, we obtain F ∈ H 2 (Ω) such that T (F ) = f and kF kH 2 (Ω) ≤
Ckf kH 1.5 (∂Ω) ; thus, we see that ∆F ∈ L2 (Ω) and that, in fact,
Z Z
DU · Dv dx = ∆F v dx ∀v ∈ H01 (Ω) . (2.62)
Ω Ω
We first establish interior regularity. Choose any (nonempty) open sets

Ω1 ⊂⊂ Ω2 ⊂⊂ Ω and let ζ ∈ C0∞ (Ω2 ) with 0 ≤ ζ ≤ 1 and ζ = 1 on Ω1 . Let
0 = min dist(spt(ζ), ∂Ω2 )/2. For all 0 < < 0 , define U (x) = η ∗ U (x)
for all x ∈ Ω2 , and set
v = −η ∗ (ζ 2 U ,j ),j .
Then v ∈ H01 (Ω) and can be used as a test function in (2.62); thus,
Z Z
− U,i η ∗ (ζ U ,j ),ji dx = − U,i η ∗ [ζ 2 U ,ij +2ζζ,i U ,j ],j dx
2
Ω Ω
Z Z
= ζ U ,ij U ,ij dx − 2 η ∗ [ζζ,i U ,j ],j U,i dx ,
2
Ω2 Ω
and
Z Z Z
2
∆F v dx = − ∆F η ∗ (ζ U ,j ),j dx = − ∆F η ∗ [ζ 2 U ,jj +2ζζ,j U ,j ] dx .
Ω Ω2 Ω2
By Young’s inequality (Theorem 2.53),
kη ∗ [ζ 2 U ,jj +2ζζ,j U ,j ]kL2 (Ω2 ) ≤ kζ 2 U ,jj +2ζζ,j U ,j kL2 (Ω2 ) ;
hence, by the Cauchy-Young inequality with δ, Lemma 2.52, for δ > 0,

Z
∆F v dx ≤ δkζD2 U k2L2 (Ω2 ) + Cδ [kDU k2L2 (Ω2 ) + k∆F k2L2 (Ω) ] .
Ω
Similarly,
Z
2 η ∗ [ζζ,i U ,j ],j U,i dx ≤ δkζD2 U k2L2 (Ω2 ) + Cδ [kDU k2L2 (Ω2 ) + k∆F k2L2 (Ω) ] .
Ω
By choosing δ < 1 and readjusting the constant Cδ , we see that
kD2 U k2L2 (Ω1 ) ≤ kζD2 U k2L2 (Ω2 ) ≤ Cδ [kDU k2L2 (Ω2 ) + k∆F k2L2 (Ω) ]
≤ Cδ k∆F k2L2 (Ω) ,
the last inequality following from (2.61), and Young’s inequality.

Since the right-hand side does not depend on > 0, there exists a sub-
sequence
0
D2 U * W in L2 (Ω1 ) .
By Theorem 2.78, U → U in H 1 (Ω1 ), so that W = D2 U on Ω1 . As weak

convergence is lower semi-continuous, kD2 U kL2 (Ω1 ) ≤ C k∆F kL2 (Ω) . As Ω1
2 (Ω) and that
and Ω2 are arbitrary, we have established that U ∈ Hloc
kU kH 2 ≤ Ck∆F kL2 (Ω) .

loc (Ω)
For any w ∈ H01 (Ω), set v = ζw in (2.62). Since u ∈ Hloc

2 (Ω), we may
integrate by parts to find that

Z
(−∆U − ∆F ) ζw dx = 0 ∀w ∈ H01 (Ω) .
Ω
Since w is arbitrary, and the spt(ζ) can be chosen arbitrarily close to ∂Ω, it
follows that for all x in the interior of Ω, we have that
−∆U (x) = ∆F (x) for almost every x ∈ Ω . (2.63)
We proceed to establish the regularity of U all the way to the boundary

∂Ω. Let {Ul }K
l=1 denote an open cover of Ω which intersects the boundary
∂Ω, and let {θl }K
l=1 denote a collection of charts such that
θl : B(0, rl ) → Ul is a C ∞ diffeomorphism ,
det Dθl = 1 ,
θl (B(0, rl ) ∩ {xn = 0}) → Ul ∩ ∂Ω ,
θl (B(0, rl ) ∩ {xn > 0}) → Ul ∩ Ω .
Let 0 ≤ ζl ≤ 1 in C0∞ (Ul ) denote a partition of unity subordinate to the

open covering Ul , and define the horizontal convolution operator, smoothing
functions defined on Rn in the first 1, ..., n − 1 directions, as follows:
Z
ρ ∗h F (xh , xn ) = ρ (xh − yh )F (yh , xn )dyh ,
Rn−1
where ρ (xh ) = −(n−1) ρ(xh /), ρ the standard mollifier on Rn−1 , and xh =
(x1 , ..., xn−1 ). Let α range from 1 to n − 1, and substitute the test function
v = − ρ ∗h [(ζl ◦ θl )2 ρ ∗h (U ◦ θl ),α ],α ◦ θl−1 ∈ H01 (Ω)

into (2.62), and use the change of variables formula to obtain the identity
Z Z
k j
Ai (U ◦ θl ),k Ai (v ◦ θl ),j dx = (∆F ) ◦ θl v ◦ θl dx , (2.64)
B+ (0,rl ) B+ (0,rl )
where the C ∞ matrix A(x) = [Dθl (x)]−1 and B+ (0, rl ) = B(0, rl )∩{xn > 0}.
We define
U l = U ◦θl , and denote the horizontal convolution operator by H = ρ ∗h .
Then, with ξl = ζl ◦ θl , we can rewrite the test function as
v ◦ θl = −H [ξl2 H U l ,α ],α .
Since differentiation commutes with convolution, we have that
(v ◦ θl ),j = −H (ξl2 H U l ,jα ),α −2H (ξl ξl ,j H U l ,α ),α ,
and we can express the left-hand side of (5.54) as

Z
Aki (U ◦ θl ),k Aji (v ◦ θl ),j dx = I1 + I2 ,
B+ (0,rl )
where
Z
I1 = − Aji Aki U l ,k H (ξl2 H U l ,jα ),α dx ,
B (0,rl )
Z+
I2 = −2 Aji Aki U l ,k H (ξl ξl ,j H U l ,α ),α dx .
B+ (0,rl )
Next, we see that

Z
I1 = [H (Aji Aki U l ,k )],α (ξl2 H U l ,jα ) dx = I1a + I1b ,
B+ (0,rl )
where
Z
I1a = (Aji Aki H U l ,k ),α ξl2 H U l ,jα dx ,
B+ (0,rl )
Z
I1b = ([H , Aji Aki ]U l ,k ),α ξl2 H U l ,jα dx ,
B+ (0,rl )
and where
[H , Aji Aki ]U l ,k = H (Aji Aki U l ,k ) − Aji Aki H U l ,k (2.65)
denotes the commutator of the horizontal convolution operator and multi-

plication. The integral I1a produces the positive sign-definite term which
will allow us to build the global regularity of U , as well as an error term:
Z
I1a = [ξl2 Aji Aki H U l ,kα H U l ,jα +(Aji Aki ),α H U l ,k ξl2 H U l ,jα ] dx ;
B+ (0,rl )
thus, together with the right hand-side of (5.54), we see that

Z Z

ξl2 Aji Aki H U l ,kα H U l ,jα dx ≤ (Aji Aki ),α H U l ,k ξl2 H U l ,jα ] dx

B+ (0,rl ) B+ (0,rl )
Z

+ |I1b | + |I2 | + (∆F ) ◦ θl v ◦ θl dx .

B+ (0,rl )
Since each θl is a C ∞ diffeomorphism, it follows that the matrix A AT is

positive definite: there exists λ > 0 such that
λ|Y |2 ≤ Aji Aki Yj Yk ∀Y ∈ Rn .
It follows that
Z Z

λ ¯
ξl2 |∂DH l 2
U | dx ≤

(Aji Aki ),α H U l ,k ξl2 H U l ,jα ] dx

B+ (0,rl ) B+ (0,rl )
Z

+ |I1b | + |I2 | + (∆F ) ◦ θl v ◦ θl dx ,

B+ (0,rl )
where D = (∂x1 , ..., ∂xn ) and p̄ = (∂x1 , ..., ∂xn−1 ). Application of the Cauchy-
Young inequality with δ > 0 shows that
Z Z

j k l 2 l
(A A ),α H U ,k ξl H U ,jα ] dx + |I2 | + (∆F ) ◦ θl v ◦ θl dx

B+ (0,rl ) i i

B+ (0,rl )
Z
≤δ ¯
ξl2 |∂DH l 2 2
U | dx + Cδ k∆F kL2 (Ω) .
B+ (0,rl )
It remains to establish such an upper bound for |I1b |.

To do so, we first establish a pointwise bound for (5.45): for Ajk = Aji Aki ,
Z
j k l
[H , Ai Ai ]U ,k (x) = ρ (xh − yh )[Ajk (yh , xn ) − Ajk (xh , xn )]U l ,k (yh , xn ) dyh
B(xh ,)
By Morrey’s inequality, |[Ajk (yh , xn ) − Ajk (xh , xn )]| ≤ CkAkW 1,∞ (B+ (0,rl )) .
Since
x − h − yh

1
∂xα ρ (xh − yh ) = 2 ρ0 ,

we see that
x − h − yh

1 0
Z
j k l
∂xα [H , Ai Ai ]U ,k (x) ≤ C ρ |U l ,k (yh , xn )| dyh

B(xh ,)
and hence by Young’s inquality,

j
∂xα [H , Ai Aki ]U l ,k 2 ≤ CkU kH 1 (Ω) ≤ Ck∆F kL2 (Ω) .

L (B+ (0,rl )
It follows from the Cauchy-Young inequality with δ > 0 that

Z
|I1b | ≤ δ ¯
ξl2 |∂DH l 2 2
U | dx + Cδ k∆F kL2 (Ω) .
B+ (0,rl )
By choosing 2δ < λ, we obtain the estimate

Z
¯
ξl2 |∂DH l 2 2
U | dx ≤ Cδ k∆F kL2 (Ω) .
B+ (0,rl )
Since the right hand-side is independent of , we find that

Z
¯
ξl2 |∂DU l 2
| dx ≤ Cδ k∆F k2L2 (Ω) . (2.66)
B+ (0,rl )
From (5.53), we know that ∆U (x) = ∆F (x) for a.e. x ∈ Ul . By the

chain-rule this means that almost everywhere in B+ (0, rl ),
−Ajk U l ,kj = Ajk ,j U l ,k +∆F ◦ θl ,
or equivalently,
−Ann Unn
l
= Ajα U l ,αj +Aβk U l ,kβ +Ajk ,j U l ,k +∆F ◦ θl .
Since Ann > 0, it follows from (5.55) that

Z
ξl2 |D2 U l |2 dx ≤ Cδ k∆F k2L2 (Ω) . (2.67)
B+ (0,rl )
Summing over l from 1 to K and combining with our interior estimates,

we have that
kukH 2 (Ω) ≤ Ck∆F kL2 (Ω) .
Step 3. k ≥ 3. At this stage, we have obtained a pointwise solution

U ∈ H 2 (Ω) ∩ H01 (Ω) to ∆U = ∆F in Ω, and ∆F ∈ H k−1 . We differentiate
this equation r times until Dr ∆F ∈ L2 (Ω), and then repeat Step 2.
Chapter 3
Fourier Series and its

Applications
3.1 Inner product spaces and Hilbert spaces
A real/complex vector space H (that is, the associated scalar field is R/C)
is called an inner product space if to each ordered pair of x and y ∈ H, there
is associated real/complex number, so-called the inner product of x and y,
such that the following rules hold:
(1) (y, x)H = (x, y)H (the bar denotes complex conjugation);
(2) (x + y, z)H = (x, z)H + (y, z)H ;
(3) (αx, y)H = α(x, y)H if x and y ∈ H and α is a scalar;
(4) (x, x)H ≥ 0 for all x ∈ H;
(5) (x, x)H = 0 only if x = 0.
We define kxkH , the norm of the vector x, to be the non-negative square

root of (x, x). Thus kxk2H = (x, x).
Theorem 3.1 (The Schwarz inequality). For x and y ∈ H,
|(x, y)H | ≤ kxkH kykH .
Proof. Put A = kxk2H , B = |(x, y)H |, and C = kyk2H . There is a complex
119
120 CHAPTER 3. Fourier Series and its Applications
number α such that |α| = 1 and α(y, x)H = B. For any real r, we have
kx − rαyk2H = (x − rαy, x − rαy)H

= kxk2H − rα(y, x)H − rᾱ(x, y)H + r2 kyk2H
= A − 2Br + Cr2 .
Since the left-hand side is non-negative, we must have B 2 ≤ AC which gives

the desired inequality.
Theorem 3.2 (Triangle inequality). For x and y ∈ H,
kx + ykH ≤ kxkH + kykH .
Proof. By the Schwarz inequality,
kx + yk2H = (x + y, x + y)H = kxk2H + (x, y)H + (y, x)H + kyk2H

2
≤ kxk2H + 2kxkH kykH + kyk2H = kxkH + kykH .

Therefore, H, k · kH is a normed space (hence a metric space as well).
If (x, y)H = 0 for some x and y ∈ H, we say that x is orthogonal to y, and
sometimes write x ⊥ y. Since (x, y)H = 0 implies (y, x)H = 0, the relation
⊥ is symmetric.
If this space is complete, then H is called a Hilbert space. From now on,
we assume that H is a Hilbert space.
Theorem 3.3 (Riesz representation). If L is a continuous linear functional

on a Hilbert space H, then there is a unique y ∈ H such that
Lx = (x, y)H ∀x ∈ H.
Proof. Without loss of generality, we may assume that Lx 6= 0 for some

x ∈ H. Then ker L ( H is a closed subspace of H; hence there exists
z ∈ (ker L)⊥ with kzkH = 1. Let u = (Lx)z − (Lz)x. Then u ∈ ker L since
Lu = (Lx)(Lz) − (Lz)(Lx) = 0 . So (u, z)H = 0, or
0 = Lx(z, z)H − Lz(x, z)H ⇒ Lx = Lz(x, z)H .
Then y = (Lz)z is what we are looking for. The uniqueness of y follows

from that if (x, y) = 0 for all x ∈ H, then y = 0.
3.1. INNER PRODUCT SPACES AND HILBERT SPACES 121
Definition 3.4 (Orthonormal sets). A set of vectors uα in a Hilbert space

H, where α runs through some index set I, is called orthonormal if it satisfies
the orthogonality relations (uα , uβ )H = 0 for all α 6= β, α ∈ I and β ∈ I,
and if it is normalized so that kuα kH = 1 for each α ∈ I. An orthonormal
set {uα }α∈I is called maximal if there is no other non-zero unit vector x ∈ H
so that (x, uα )H = 0 for all α ∈ I. In other words, {uα }α∈I is maximal if
(x, uα )H = 0 for all α ∈ I implies x = 0.
Definition 3.5 (Fourier coefficients). If {uα }α∈I is orthonormal, we asso-

ciate with each x ∈ H a real/complex function x̂ on the index set I, defined
by
x̂(α) = (x, uα )H α∈I.
The number x̂(α) sometimes is called the Fourier coefficients of x, relative

to the set {uα }α∈I .
Theorem 3.6. Suppose that {uα }α∈I is an orthonormal set in H and that
F is a finite subset of I. Let MF be the span of {uα }α∈F .
(1) If ϕ is a complex function on I that is 0 outside F , then there is a

vector y ∈ MF , namely
X
y= ϕ(α)uα
α∈F
X
that has ŷ(α) = ϕ(α) for every α ∈ I. Also, kyk2H = |ϕ(α)|2 .
α∈F
P
(2) If x ∈ H and sF (x) = x̂(α)uα , then
α∈F
kx − sF (x)kH ≤ kx − skH ∀ s ∈ MF ,
and
X
|x̂(α)|2 ≤ kxk2H . (Bessel’s inequality)
α∈F
Theorem 3.7. If {uα }α∈I is an orthonormal subset of a Hilbert space H,

then the following conditions are equivalent:
(1) (x, uα )H = 0 for all α ∈ I implies that x = 0;

P
(2) x = (x, uα )H uα for all x ∈ H;
α∈I
P 2
(3) kxk2H = (x, uα )H for all x ∈ H;
α∈I
(4) The smallest closed linear subspace that contains {uα }α∈I is H;
(5) {uα }α∈I is a maximal orthonormal set.
3.2 The Hilbert space L2 (T)
A 2π-periodic function on R may be identified with a function on the circle,

or one-dimensional torus, T = R/(2πZ), which we define by identifying
points in R that differ by 2πn for some n ∈ Z. We use C(T) to denote the
space of continuous functions on R with period 2π. The space L2 (T) is the
completion of C(T) with respect to the L2 -norm
hZ i1
2
kf kL2 (T) = |f (x)|2 dx .
T
The space L2 (T) is a Hilbert space with the inner product

Z
(f, g)L2 (T) = f (x)g(x)dx .
T
The Fourier basis elements are the functions

1
ek (x) = √ eikx .
2π
The set {ek }∞ 2

k=−∞ is an orthonormal set in L (T). The goal in this section
is to show that {ek }∞ 2
k=−∞ is maximal. In other words, for each f ∈ L (T),
∞ ∞ Z
X 1 X
f= (f, ek )L2 (T) ek = f (y)eik(x−y) dy .
2π T
k=−∞ k=−∞
3.2.1 Approximations of the identity

Definition 3.8. A family of functions ϕn ∈ C(T) n ∈ N is an approxi-
mation of the identity if
(1) ϕn (x) ≥ 0;
§3.2 The Hilbert space L2 (T) 123
Z
(2) ϕn (x)dx = 1 for every n ∈ N;
T
Z
(3) lim ϕn (x)dx = 0 for every δ > 0, here we identify T with the
n→∞ δ≤|x|≤π
interval [−π, π].
Definition 3.9 (Convolutions on T). The convolution of two continuous

function f, g : T → C is the continuous function f ∗ g : T → C defined by
the integral
Z
(f ∗ g)(x) = f (x − y)g(y) dy .
T
Note that all the conclusions from Section A.7 are still valid. In partic-
ular, we have
Theorem 3.10. If {ϕn }∞

n=1 is an approximation of the identity and f ∈
C(T), then ϕn ∗ f converges uniformly to f as n → ∞.
Proof. Without loss of generality, we may assume that f 6≡ 0. By the

definition of the convolution,
Z

(ϕn ∗ f )(x) − f (x) = ϕn (x − y)f (y)dy − f (x)
ZT

= ϕn (x − y) f (x) − f (y) dy ,
T
where we use (2) of Definition 3.8 to obtain the last equality. Now given

> 0. Since f ∈ C(T), there exists δ > 0 such that |f (x) − f (y)| <
2
whenever |x − y| < δ. Therefore,
|(ϕn ∗ f )(x) − f (x)|

Z Z

≤ ϕn (x − y) f (x) − f (y) dy +
ϕn (x − y)f (x) − f (y)dy
|x−y|<δ δ≤|x−y|
Z Z
≤ ϕn (x − y)dy + 2 max |f | ϕn (z)dz .
T T δ≤|z|≤π
By (3) of Definition 3.8, there exists N > 0 such that if n ≥ N ,

Z
ϕn (z)dx ≤ .
δ≤|z|≤π 4 maxT |f |

Therefore, for n ≥ N , (ϕn ∗ f )(x) − f (x) ≤ . for all x ∈ T.
3.2.2 Trigonometric polynomials
Definition 3.11. A trigonometric polynomial of degree n is a finite sum of

the form
n
c0 X
f (x) = + (ck cos nx + sk sin nx) x ∈ R.
2
k=1
The collection of all trigonometric polynomial of degree n is denoted by

Pn (T).
On account of the Euler identity eiθ = cos θ+i sin θ, the expression above
can also be written in the form
n
X ck − isk
f (x) = ak eikx with ak = ,
2
k=−n
where s0 is taken to be 0.
Theorem 3.12. The trigonometric polynomials are dense in C(T) with

respect to the uniform norm.
Z
Proof. Let ϕn (x) = cn (1+cos x)n , where cn is chosen so that ϕn (x)dx = 1.
T
By the residue theorem,
(z + 1)2n

z + 1 n dz 1 π 2n
Z I I
(1 + cos x)n dx = 1+ = n n+1
dz = n−1
;
T S1 2z iz i2 S1 z 2 n
2n−1 (n!)2
thus cn = .
π (2n)!
Now {ϕn }∞
n=1 is clearly non-negative and satisfies (2) of Definition 3.8
for all n ∈ N. Let δ > 0 be given.
Z Z 1 + cos δ n (n!)2
ϕn (x)dx ≤ cn (1 + cos δ)n dx ≤ 22n .
δ≤|x|≤π δ≤|x|≤π 2 (2n)!
n!
By Stirling’s formula lim √ = 1,
n→∞2πnnn e−n
√ 2
2n 1 + cos δ 2πnnn e−n
Z n
lim ϕn (x)dx ≤ lim 2 p
n→∞ δ≤|x|≤π n→∞ 2 2π(2n)(2n)2n e−2n
√ 1 + cos δ n
= lim πn = 0.
n→∞ 2
So {ϕn }∞
n=1 is an approximation of the identity. By Theorem 3.10, ϕk ∗ f
converges uniformly to f if f ∈ C(T), while ϕn ∗f is a trigonometric function.

§3.2 The Hilbert space L2 (T) 125
Corollary 3.13. Given f ∈ L2 (T), define

Z π n
1 1 X ˆ
fˆ(k) = √ f (x)e−ikx dx , and sn (f, x) = √ f (k)eikx .
2π −π 2π k=−n
Then lim ksn (f, ·) − f kL2 (T) = 0, and

n→∞
Z ∞
|fˆ(k)|2 .
X
|f (x)|2 dx = (Parseval’s identity) (3.1)
T k=−∞
Proof. By Theorem 3.7 (4), it suffices to show that the smallest closed sub-
space containing {ek }∞ 2
k=−∞ is L (T). However, since every linear combina-
tion of ek can be expressed as a trigonometric polynomial, and by Theorem
3.12 the space of trigonometric polynomials is dense in C(T), we know that
the space spanned by {ek }∞
k=−∞ is dense in C(T). The implication from
uniform convergence to L2 (T)-convergence then guarantees that the space
spanned by {ek }∞ 2
k=−∞ is dense in L (T).
Remark 3.14. Using the Euler identity, we can express sn (f, x) defined as
n
c0 X
sn (f, x) = + ck cos kx + sk sin kx ,
2
k=1
where
π π
1 1
Z Z
ck = f (x) cos kxdx , sk = f (x) sin kxdx .
π −π π −π
The corresponding Parseval’s identity is

h ∞
X i
kf k2L2 (T) = π c20 + (c2k + s2k ) .
k=1
The proof of the following lemma is left as an exercise.
Lemma 3.15. Let f, g ∈ L2 (T). Then

∞
fˆ(k)ĝ(k).
X
(f, g)L2 (T) =
k=−∞
By Parseval’s identity (3.1), for f ∈ L2 (T),
lim |fˆ(k)| = 0 .
|k|→∞
In fact, the Fourier coefficient of a function f ∈ L1 (T) also converges to 0

which is the Riemann-Lebesgue Lemma.
Lemma 3.16 (Riemann-Lebesgue). For f ∈ L1 (T), fˆ(k) → 0 as |k| → ∞.
Proof. Let f ∈ L1 (T). Given > 0, there exists f ∈ L2 (T) such that

kf − f kL1 (T) < . For this , there exists N > 0 such that |fˆ (k)| <
2 2
whenever k ≥ N . Therefore, for k ≥ N ,
|fˆ(k)| ≤ |fˆ(k) − fˆ (k)| + |fˆ (k)|

Z
f (x) − f (x) ek (x)dx + |fˆ (k)|

=
T
≤ kf − f kL1 (T) + |fˆ (k)| <
which implies that lim fˆ(k) = 0.

|k|→∞
Definition 3.17 (Weak convergence). A sequence un ∈ H is said to con-

verge weakly to u ∈ H if
lim (un , g)H = (u, g)H ∀g ∈ H.

n→∞
We use the notation un * u in H to denote the weak convergence of un to

u in H.
With this definition, by the Riemann-Lebesgue Lemma we have the fol-

lowing
Theorem 3.18. The Fourier basis en converges weakly to 0 in L2 (T).
3.2.3 Fourier representation of functions on [0, π]
Any functions defined on [0, π] can be viewed as the restriction of an even/odd

function defined on [−π, π] to [0, π]. An even/odd function f in L2 (T) can
be expressed as
∞ ∞
c0 X .X
f (x) = + ck cos kx sk sin kx .
2
k=1 k=1
For a function f ∈ L2 (0, 2π) (here we identify T with [0, 2π]), g(x) = f (2x)
is a function in L2 (0, π). Since
∞
c0 X
f (x) = + (ck cos kx + sk sin kx) ,
2
k=1
3.3. POINTWISE AND UNIFORM CONVERGENCE OF THE FOURIER SERIES127
we have
∞
c0 X
g(x) = + (ck cos 2kx + sk sin 2kx) ,
2
k=1
where
2 2
Z Z
ck = g(x) cos kxdx , sk =
g(x) sin kxdx .
T π π T
n 1 r2 o∞ nr 2 o∞
As a consequence, √ , cos kx , sin kx are both maxi-
π π k=1
rπ k=1
r
n 1 2 2 o∞
mal orthonormal sets on L2 (0, π). So is √ , cos 2kx, sin 2kx .
r π π π k=1
n 1 2 o∞
We note that ± √ , ± cos kx is the collection of all non-trivial
π π k=1
functions with unit L2 (0, π)-norm satisfying
uxx = λu for some λ ∈ R ,

ux (0) = ux (π) = 0 ,
r
n 2 o∞
while ± sin kx is the collection of all non-trivial functions with
π k=1
2
unit L (0, π)-norm satisfying
uxx = λu for some λ ∈ R ,

u(0) = u(π) = 0 .
3.3 Pointwise and uniform convergence of the Fourier

series
Given f ∈ L2 (T), by Corollary 3.13 we know that sn (f, ·) → f in L2 (T); thus

possesses a subsequence snj (f, ·) which converges to f almost everywhere. In
this section, instead of assuming that f ∈ L2 (T), we consider f ∈ C 0,α (T),
and investigate the behavior of convergence of the Fourier series of f . We
first show that sn (f, ·) converges to f pointwise (p.w.) and then show that
the convergence is in fact uniform.
Before proceeding, we define
n
1 e−inx ei(2n+1)x − 1

1 X ikx
Dn (x) = e =
2π 2π eix − 1
k=−n
1 ei(n+1/2)x − e−i(n+1/2)x sin(n + 12 )x

= = .
2π eix/2 − e−ix/2 2π sin x2
Then
n n
1
Z
fˆ(k)eikx =
X X
sn (f, x) = f (y)eik(x−y) dy
2π T
k=−n k=−n
Z
= f (y)Dn (x − y)dy = (Dn ∗ f )(x) ,
T
and Corollary 3.13 states that Dn ∗ f → f in L2 (T) for all f ∈ L2 (T).
Definition 3.19. The function
sin(n + 12 )x
Dn (x) = (3.2)
2π sin x2
is called the Dirichlet kernel.
3.3.1 Pointwise convergence

Z
By the construction of the Dirichlet kernel Dn , Dn (x)dx = 1 for all
T
n ∈ N. However, as one can see from (3.2), Dn attains negative values in T,
so Dn fails to be an approximation of the identity. Therefore, at first glance
we would not expect that sn (f, ·) ≡ Dn ∗ f converges to f uniformly because
one cannot apply Theorem 3.10; thus we focus on achieving the pointwise
convergence first.
Without loss of generality, we show that sn (f, 0) → 0 as n → ∞ for
otherwise we may consider g(x) = f (x+a)−f (a) and show that sn (g, 0) → 0
which implies that sn (f, a) → f (a).
By definition,
1 π sin(n + 21 )x
Z
sn (f, 0) = (Dn ∗ f )(0) = f (x) dx (3.3)
π −π 2 sin x2
1 π sin nx cos x2 1 π
Z Z
= f (x) dx + f (x) cos nxdx ;
π −π 2 sin x2 π −π
thus
1 π sin nx
Z
sn (f, 0) − f (x) dx
π −π x
1 π
Z cos x 1 1 π
Z
2
= f (x) − sin nxdx + f (x) cos nxdx .
π −π 2 sin x2 x π −π
§3.3 Pointwise and uniform convergence of Fourier series 129
cos x2 1
Note that the function g(x) ≡ − can be made continuous at x = 0
2 sin x2 x
if
cos x 1 x cos x2 − 2 sin x2
2
g(0) ≡ lim − = lim = 0.
x→0 2 sin x x x→0 2x sin x2
2
Therefore, by Riemann-Lebesgue Lemma,

1 π sin nx
Z
sn (f, 0) − f (x) dx → 0 as n → ∞.
π −π x
However, since f ∈ C 0,α (T) and f (0) = 0, the function f (x)/x ∈ L1 (T);
thus by Riemann-Lebesgue Lemma again,
1 π sin nx
Z
f (x) dx → 0 as n→∞
π −π x
which implies that sn (f, 0) → 0 as n → ∞. This proves the pointwise

convergence of the Fourier series of a Hölder continuous function.
3.3.2 Uniform convergence
Before proceeding to the proof of uniform convergence of the Fourier series

(of Hölder continuous functions), we make the following observation. We
consider the Fejér kernel which is the Cesàro mean of the Dirichlet kernel
defined by
n (n+1)x
1 X 1 − cos(n + 1)x 1 sin2 2
Fn (x) = Dk (x) = = .
n+1
k=0
4π(n + 1) sin2 x2 2π(n + 1) sin2 x
2
Z
We note that Fn ≥ 0 has the property that Fn (x)dx = 1. Moreover, for
T
any δ > 0,
1
Z Z
lim Fn (x)dx ≤ lim δ
dx = 0 ;
n→∞ δ<|x|≤π n→∞ δ<|x|≤π 2π(n + 1) sin2 2
hence {Fn }∞
n=1 is an approximation of the identity. Define σn (f, x) = (Fn ∗
f )(x). By Theorem 3.10, σn (f, ·) converges to f uniformly as n → ∞ if
f ∈ C(T).
However, σn (f, ·) ≡ Fn ∗ f → f uniformly as n → ∞ is still far away
from our goal: Dn ∗ f → f uniformly as n → ∞. In the following, we first
consider an easier case f ∈ C 0,1 (T); that is, f is Lipschitz continuous on
T. We note that if f ∈ C 0,1 (T), then f is absolutely continuous; thus is

differentiable a.e. and satisfies the integration by parts formula
Z b x=b Z b
0
f (x)g (x)dx = f (x)g(x) − f 0 (x)g(x)dx ∀ g ∈ C 1 (T) .

a x=a a
The additional differentiability (or more precisely, the validity of the integra-
tion by parts formula) of Lipschitz continuous functions allows us to prove
the uniform convergence more easily. We have the following
Theorem 3.20. Suppose that f ∈ C 0,1 (T). Then Dn ∗ f converges to f

uniformly as n → ∞.
Z
Proof. Since Dn (x − y)dy = 1 for all x ∈ T,
T
Z

sn (f, x) − f (x) = (Dn ∗ f − f )(x) = Dn (x − y) f (y) − f (x) dy
Z T

= Dn (y) f (x + y) − f (x) dy .
T
Let L ≡ n + 1/2, then the goal is to show that
Z f (x + y) − f (x)
lim sup sin Ly dy = 0 .

L→∞ x∈T T sin y2
We break the integral into two parts, one integrating over |y| ≤ δ and
one integrating over δ < |y| ≤ π. Since f ∈ C 1 (T), |f (y + x) − f (x)| ≤
kf 0 kL∞ (T) |y|; thus
Z δ
Z f (x + y) − f (x) 0 y
sin Ly dy ≤ kf kL∞ (T) y dy ≤ Cδ . (3.4)

y
sin 2 −δ sin 2

|y|≤δ
As for the integral over δ < |y| ≤ π, we note that

Z π
f (x + y) − f (x) cos Ly f (x + y) − f (x) y=π
sin Ly y dy = −
sin 2 L sin y2

δ y=δ
Z π
cos Ly d f (x + y) − f (x)
+ dy .
δ L dy sin y2
Note that the first term on the right-hand side converges to 0 uniformly as
L → ∞. For the second term,
Z π cos Ly d f (x + y) − f (x)
dy

y
L dy sin 2

δ
Z π cos Ly f 0 (x + y) Z π cos Ly cos y f (x + y) − f (x)
2
≤ dy + dy

δ L sin y2 δ L sin2 y2
π−δ π−δ π−δ
≤ kf 0 kL∞ (T) δ
+ kf kL∞ (T) 2 δ
≤ kf kC 0,1 (T) .
L sin 2 L sin 2 L sin2 2δ
Therefore, combining the estimate above with (3.4), we find that

Z f (x + y) − f (x)
lim sup sup sin Ly dy ≤ Cδ ;

L→∞ x∈T T sin y2
and the conclusion follows from that δ > 0 is chosen arbitrarily.
The uniform convergence of sn (f, ·) to f for f ∈ C 0,α (T) with α ∈ (0, 1)

requires a lot more work. The idea is to estimate f −sn (f, ·) L∞ (T) in terms
of the quantity inf kf − pkL∞ (T) . Since sn (f, ·) ∈ Pn (T), it is obvious
p∈Pn (T)
that

inf kf − pkL∞ (T) ≤ f − sn (f, ·) L∞ (T) .
p∈Pn (T)
The goal is to show the inverse inequality

f − sn (f, ·) ∞ ≤ Cn inf kf − pkL∞ (T) (3.5)
L (T) p∈Pn (T)
for some constant Cn , and pick a suitable p ∈ Pn (T) which gives a good

upper bound for f − sn (f, ·) L∞ (T) . The inverse inequality is established
via the following
Proposition 3.21. The Dirichlet kernel Dn satisfies that for all n ∈ N,

Z

Dn (x)dx ≤ 2 + log n . (3.6)
T
Proof. The validity of (3.6) for the case n = 1 is left to the reader, and we
n
X eikx
provide the proof for the case n ≥ 2 here. Recall that Dn (x) = =
2π
k=−n
sin(n + 21 )x
. Therefore,
2π sin x2
1 sin(n + 1 )x
Z Z π Z
n
Z π
2

Dn (x)dx = 2 Dn (x)dx = 2Dn (x)dx +

dx .

π sin x2

1
T 0 0 n
The first integral can be estimated by

Z 1
n 1 2n + 1
2Dn (x)dx ≤ . (3.7)
0 π n
2x π
Since ≤ sin x for 0 ≤ x ≤ , the second integral can be estimated by
π 2
Z π 1 Z π
sin(n + 2 )x 1

dx ≤ dx = log π + log n . (3.8)
π sin x2 1 x

1
n n
2n + 1
We then conclude (3.6) from (3.7) and (3.8) by noting that log π + ≤
nπ
2 for all n ≥ 2.
Remark 3.22. A more subtle estimate can be done to show that

Z

Dn (x)dx ≥ c1 + c2 log n ∀n ∈ N
T
for some positive constants c1 and c2 .
With the help of Proposition 3.21, we are able to prove the inverse in-
equality (3.5). The following theorem is a direct consequence of Proposition
3.21.
Theorem 3.23. Let f be a 2π-periodic function. Then

f − sn (f, ·) ∞ ≤ (3 + log n) inf kf − pkL∞ (T) . (3.9)
L (T) p∈Pn (T)
Proof. For n ∈ N and x ∈ T,

Z

sn (f, x) ≤ Dn (y)|f (x − y)dy ≤ (2 + log n)kf kL∞ (T) .
T
Given > 0, let p ∈ Pn (T) such that
kf − pkL∞ (T) ≤ inf kf − pkL∞ (T) + .

p∈Pn (T)
Then by the fact that sn (p, x) = p(x) if p ∈ Pn (T), we obtain that

f − sn (f, ·) ∞ ≤ f − p ∞ + p − sn (f, ·) ∞
L (T) L (T) L (T)

≤ f − p L∞ (T) + sn (f − p, ·) L∞ (T)

≤ f − p L∞ (T) + (2 + log n)kf − pkL∞ (T)
h i
≤ (3 + log n) inf kf − pkL∞ (T) + ,
p∈Pn (T)
and (3.9) is obtained by passing → 0.
Having established Theorem 3.23, the study of the uniform convergence

of sn (f, ·) to f then amounts to the study of the quantity inf kf −
p∈Pn (T)
pkL∞ (T) . In Exercise Problem 3.2, the reader is asked to show that
1 + 2 log n
inf kf − pkL∞ (T) ≤ kf kC 0,1 (T) ;
p∈Pn (T) 2n
thus by Theorem 3.23, sn (f, ·) converges to f uniformly as n → ∞ if f ∈

C 0,1 (T), a restatement of Theorem 3.20.
The estimate of inf kf − pkL∞ (T) for f ∈ C 0,α (T) is more difficult,
p∈Pn (T)
and requires a clever choice of p. We begin with the following
Lemma 3.24. If f is a continuous function on [a, b], then for all δ1 > 0,
f (x) − f (y) ≤ 1 + δ1

sup sup f (x) − f (y) .
|x−y|≤δ1 δ2 |x−y|≤δ2
The proof of Lemma 3.24 is not very difficult, and is left to the readers.
Now we are in the position of prove the theorem due to D Jackson.
Theorem 3.25 (Jackson). Let f be a 2π-periodic continuous function.

Then for some constant C > 0,
inf kf − pkL∞ (T) ≤ C sup |f (x) − f (y)| .

p∈Pn (T) 1
|x−y|≤ n
Proof. Let p(x) = 1 + c1 cos x + · · · + cn cos nx be a positive trigonometric

function of degree n with coefficients {ci }ni=1 determined later. Define an
operator K on C(T) by
1
Z
Kf (x) = p(y)f (x − y)dy .
2π T
Then Kf ∈ Pn (T). Lemma 3.24 then implies
Kf (x) − f (x) ≤ 1
Z

p(y)f (x − y) − f (x)dy
2π T
Z π
1
≤ p(y) 1 + n|y| sup f (x) − f (y)dy
2π −π |x−y|≤ n1
π
n
h Z i
= 1+ |y|p(y)dy sup f (x) − f (y) .
2π −π |x−y|≤ 1 n
π2
By Hölder’s inequality and that y 2 ≤ (1 − cos y) for y ∈ T,
2
Z π h 1 Z π i1 h 1 Z π i1
1 2 2
|y|p(y)dy ≤ y 2 p(y)dy p(y)dy
2π −π 2π −π 2π −π
hπ Z π i1 π√
2
≤ (1 − cos y)p(y)dy = 2 − c1 .
4 −π 2
Therefore,
nπ √
kKf − f kL∞ (T) ≤ 1 + 2 − c1 sup f (x) − f (y) .
2 |x−y|≤ 1 n
√
To conclude the theorem, we need to show that the number n 2 − c1 can
be made bounded by choosing p properly. Nevertheless, let
n n X
n
X (k + 1)π ikx 2 X (k + 1)π (` + 1)π i(k−`)x
p(x) = c sin e =c sin e

n+2 n+2 n+2
k=0 k=0 `=0
n n
X (k + 1)π X (k + 1)π (` + 1)π
=c sin2 + 2c sin cos(k − `)x
n+2 n+2 n+2
k=0 k, ` = 0
k 6= `
≡ 1 + c1 cos x + · · · + cn cos nx .
By comparison of coefficients of p, we find that

n n
−1
X
2 (k + 1)π 1 Xh 2(k + 1)π i
c = sin = 1 − cos
n+2 2 n+2
k=0 k=0
(2n+3)π π
n + 1 sin n+2 − sin n+2 n+2
= − π = ,
2 4 sin n+2 2
and
n n
X (k + 1)π kπ Xh π (2k + 1)π i
c1 = 2c sin sin =c cos − cos
n+2 n+2 n+2 n+2
k=1 k=1
(2n+2)π 2π i
h π sin n+2 − sin n+2
= c n cos − π
n+2 2 sin n+2
2π i
h π sin n+2
= c n cos + π
n + 2 sin n+2
π π
= c(n + 2) cos = 2 cos .
n+2 n+2
Therefore,
√ π 12 π
n 2 − c1 = n 2 − 2 cos = 2n sin
n+2 2(n + 2)
π π
= 2(n + 2) sin − 4 sin
2(n + 2) 2(n + 2)
2(n + 2) π π
=π sin − 4 sin
π 2(n + 2) 2(n + 2)
which is bounded by π + 4.
Finally, since lim n−α log n = 0 for all α > 0 , we conclude the following
n→∞
Theorem 3.26. Suppose that f ∈ C 0,α (T) for some α ∈ (0, 1). Then Dn ∗f
converges to f uniformly as n → ∞.
Remark 3.27. The converse of Theorem 3.25 is the Bernstein theorem

which states that if f is a 2π-periodic function such that for some constant
C (independent of n) and α ∈ (0, 1),
inf kf − pkL∞ (T) ≤ Cn−α (3.10)

p∈Pn (T)
for all n ∈ N, then f ∈ C 0,α (T). In other words, (3.10) is an equivalent

condition to the Hölder continuity with exponent α of 2π-periodic continu-
ous functions. One way of proving the Bernstein theorem can be found in
Exercise Problem 3.4.
3.3.3 Jump discontiunity and Gibbs phenomenon
In this section, we study the convergence of the Fourier series of functions

with jump discontinuities. We show that the Fourier series evaluated at the
jump discontinuity converges to the average of the limits from the left and
the right. Moreover, the convergence of the Fourier series is never uniform
in the domain excluding these jump discontinuities due to the famous Gibbs
phenomenon: near the jump discontinuity the maximum difference between
the limit of the Fourier series and the function itself is at least 8% of the
jump. To be more precise, we have the following
Theorem 3.28. Let f : R → R be a piecewise continuously differentiable

function which is periodic with some period L > 0. Suppose that at some
point x0 the limit from the left f (x− +
0 ) and the limit from the right f (x0 ) of
the function f exist and differ by a non-zero gap a:
−
f (x+
0 ) − f (x0 ) = a 6= 0,
then there exists a constant c > 0, independent of f , x0 and L (in fact,

1 π sin x 1
Z
c= dx − ≈ 0.089490), such that
π 0 x 2
L
lim sn f, x0 + = f (x+
0 ) + ca , (3.11a)
n→∞ 2n
L
lim sn f, x0 − = f (x−
0 ) − ca . (3.11b)
n→∞ 2n
Moreover,
−
f (x+
0 ) + f (x0 )
lim sn (f, x0 ) = . (3.12)
n→∞ 2
Proof. Without loss of generality, we may assume that x0 = 0 is the only
f (0+ ) + f (0− )
discontinuity of f , f (0) = , and L = 2π. Let g be a discon-
2
tinuous function defined by
 a
 (x + π) if −π ≤ x < 0 ,
 2π


g(x) = 0 if x = 0 ,

 a

 (x − π) if 0 < x ≤ π .
2π
Then F = f + g is Lipchitz continuous on T, thus by Theorem 3.20,
f (0+ ) + f (0− )
= F (0) = lim sn (F, 0) = lim sn (f, 0) + lim sn (g, 0)
2 n→∞ n→∞ n→∞
= lim sn (f, 0) .
n→∞
This proves (3.12). Z

π
1 ia
By ĝ(k) = √ g(x)e−ikx dx = √ if k 6= 0, and ĝ(0) = 0, we
2π −π 2πk
n n
X ĝ(k) X a
find that sn (g, x) = √ eikx = − sin(kx); thus
2π πk
k=−n k=1
n n
π a kπ aX n kπ π a π sin x
X Z
sn g, =− sin =− sin →− dx .
n πk n π kπ n n π 0 x
k=1 k=1
Therefore, by the continuity of F ,

f (0+ ) + f (0− ) π π π
= lim F = lim sn f, + lim sn g,
2 n→∞ n n→∞ n n→∞ n
π a π sin x
Z
= lim sn f, − dx ,
n→∞ n π 0 x
f (0+ ) + f (0− ) a
and (3.11a) follows from + = f (0+ ). (3.11b) can be proved
2 2
in the same fashion, and is left as an exercise.
3.4 The Sobolev space H s (T)
Definition 3.29. For s > 0 (not necessary an integer), the Sobolev space
H s (T) consists of all functions f ∈ L2 (T) such that
∞
|k|2s |fˆ(k)|2 < ∞ .
X
k=−∞
§3.4 The Sobolev space H s (T) 137
If f, g ∈ H s (T), then the H s (T)-inner product of f and g is defined by

∞
(1 + |k|2 )s fˆ(k)ĝ(k)
X
(f, g)H s (T) =
n=−∞
which induces the H s (T)-norm as

∞
(1 + |k|2 )s |fˆ(k)|2 .
X
kf k2H s (T) =
k=−∞
Example 3.30. Let the heavyside function H be defined by

1 if 0 ≤ x ≤ π ,
H(x) =
0 if −π ≤ x < 0 .
Then H ∈ L2 (T). Moreover, we have the Fourier coefficients for H :
 r
π
if k = 0,


2


Z π 
1 
Ĥ(k) = √ e−ikx dx = 0 if k is even, k 6= 0,
2π 0 

 2


if k is odd,

ik
1
hence H ∈ H s (T) if s < .
2
Theorem 3.31. Let 0 < r < t < ∞, and s = αr + (1 − α)t for some
α ∈ (0, 1). Then
kukH s (T) ≤ kukαH r (T) kuk1−α

H t (T) . (3.13)
Proof. By definition,
∞
X
kuk2H s (T) = (1 + |k|2 )s |û(k)|2
k=−∞
X∞
= (1 + |k|2 )αr |û(k)|2α (1 + |k|2 )(1−α)t |û(k)|2(1−α) .
k=−∞
1 1
Noting that −1
+ = 1, by the Hölder inequality we find that
α (1 − α)−1
∞
X
(1 + |k|2 )αr |û(k)|2α (1 + |k|2 )(1−α)t |û(k)|2(1−α)
k=−∞
∞
h X ∞
iα h X i1−α
≤ (1 + |k|2 )r |û(k)|2 (1 + |k|2 )t |û(k)|2
k=−∞ k=−∞
which leads to (3.13).

3.4.1 Characterization of H 1 (T)
Definition 3.29 gives a quantitative way of describing functions in H s (T).

In this section, a qualitative point of view of H 1 (T) is provided based on
the Hahn-Banach theorem from functional analysis. Roughly speaking, a
function f ∈ H 1 (T) has weak derivatives satisfying the integration by parts
formula. We start from stating the following
Theorem 3.32 (Hahn-Banach). If Y is a linear subspace of a normed linear

space X and T : Y → R is a bounded linear functional on Y with kT k = M ,
then there is a bounded linear functional T̃ : X → R on X such that T̃
restricted to Y is equal to T and kT̃ k = M .
In other words, a bounded linear functional on a normed linear space

can be extended to a bounded linear functional on a bigger space.
Let f ∈ H 1 (T) and ϕ ∈ C 1 (T). We define a (bounded) linear functional
Tf on C 1 (T) by
Z
Tf (ϕ) = f (x)ϕ 0 (x)dx .
T
The goal is to extend Tf to a bounded linear functional T˜f defined on L2 (T).

Since the application the Hahn-Banach theorem requires that the range of
the linear function to be real, in the following discussion we will always
assume that f and ϕ are real-valued functions.
Since ϕ 0 ∈ C(T) ⊆ L2 (T), we can compute ϕ̂ 0 and obtain that ϕ̂ 0 (k) =
ik ϕ̂(k). Therefore,
Z ∞
fˆ(k)ik ϕ̂(k) ;
X
f (x)ϕ 0 (x)dx = (f, ϕ 0 )L2 (T) =
T k=−∞
hence
Z ∞
|k||fˆ(k)||ϕ̂(k)| ≤ kf kH 1 (T) kϕkL2 (T) .
X
0
f (x)ϕ (x)dx ≤

T n=−∞
The computation above shows that if f ∈ H 1 (T), Tf is a bounded linear

functional (on a subspace of L2 (T)). By the Hahn-Banach theorem, Tf can
be extended to a bounded linear functional T˜f : L2 (T) → R. By the Riesz
§3.4 The Sobolev space H s (T) 139
representation theorem, there is a function g ∈ L2 (T) such that

Z
T̃f (ϕ) = (ϕ, g)L2 (T) = ϕ(x)g(x)dx ∀ ϕ ∈ L2 (T) .
T
In particular, for ϕ ∈ C 1 (T),

Z Z
ϕ(x)g(x)dx = T̃f (ϕ) = Tf (ϕ) = f (x)ϕ 0 (x)dx .
T T
The function h = −g is called the weak derivative of f , and usually is
denoted by f 0 as well. The reason for calling h the weak derivative of f is
that if f ∈ C 1 (T), then
Z Z
0
− f (x)ϕ(x)dx = f (x)ϕ 0 (x)dx , (3.14)
T T
so h is indeed the derivative of f . Note that g ∈ L2 (T) is “the same as”

saying that f 0 ∈ L2 (T). Therefore, the space H 1 (T) consists of all L2 (T)-
functions whose weak derivatives are also L2 (T)-functions.
Remark 3.33. H 1 (T) is in fact the completion of (C 1 (T), k · kH 1 (T) ). In

other words, let f ∈ H 1 (T), then for any given > 0, there is a function
f ∈ C 1 (T) such that
kf − f kH 1 (T) < .
Remark 3.34. The Hahn-Banach theorem does not guarantee the unique-
ness of the extension T̃f . Therefore, there might be two extensions T̃f1 and
T̃f2 mapping from L2 (T) to R that equal Tf on C 1 (T). Suppose that g1 and
g2 are the corresponding representations of T̃f1 and T̃f2 . By definition,
Z Z
ϕ(x)g1 (x)dx = T̃f1 (ϕ) = Tf (ϕ) = T̃f2 (ϕ) = ϕ(x)g2 (x)dx
T T
for all ϕ ∈ C 1 (T). Therefore, g1 = g2 a.e. in L2 (T); thus the extension T̃f
is indeed unique. The key here is that C 1 (T) is dense in L2 (T).
3.4.2 The Fourier representation of f 0 for f ∈ H 1 (T)

1
Let ϕ(x) = √ eikx in (3.14). By ϕ̂(`) = δ`k , we find that
2π
∞ Z
ˆ ˆ
X
0
f (k) = 0
f (`)ϕ̂(`) = f 0 (x)ϕ(x)dx
`=−∞ T
ik
Z Z
=− f (x)ϕ 0 (x)dx = √ f (x)e−ikx dx = ik fˆ(k) .
T 2π T
Therefore,
∞ ∞
dh 1 X ˆ i 1 X
√ f (k)eikx = f 0 (x) = √ ik fˆ(k)eikx
dx 2π k=−∞ 2π k=−∞
∞
1 X d hˆ ikx
i
= √ f (k)e .
2π k=−∞ dx
d d
In other words, if denotes the weak differentiation operator, then
dx dx
commutes with the infinite sum (in which the convergence of the infinite
sum is understood in the L2 -sense).
Theorem 3.35 (Sobolev embedding, a simple version). If f ∈ H s (T) for

s > 1/2, then there exists f˜ ∈ C(T) so that f = f˜ almost everywhere.
Moreover, there exists a constant Cs > 0 such that
kf kL∞ (T) ≤ Cs kf kH s (T) ∀ f ∈ H s (T) . (3.15)
Proof. Let sn (f, x) be the partial sum of the Fourier series of f defined as
before. Then for n ≥ m,
1 X ˆ 1
|fˆ(k)|
X
ikx
|sn (f, x) − sm (f, x)| = √ f (k)e ≤ √
2π m<|k|≤n 2π m<|k|≤n
1 h X i1/2 h X 1 i1/2
≤√ (1 + |k|2 )s |fˆ(k)|2 .
2π m<|k|≤n (1 + |k|2 )s
m<|k|≤n
Therefore, ksn (f, ·) − sm (f, ·)kL∞ (T) → 0 as n, m → ∞, which implies that

sn (f, ·) converges uniformly; hence f˜ = lim sn (f, ·) is continuous.
n→∞
∞ i1
1 h X 1 2
The constant Cs in (3.15) can be chosen as √ 2 s
.
2π (1 + |k| )
k=−∞
nr 1 r 2 o∞
Remark 3.36. As mentioned in Section 3.2, , cos kx is an
r π r π k=1
1 2 cos kx
orthonormal basis of L2 (0, π). Let w0 = and wk = √ . Then
π π 1 + k2
{wk }∞ 1
k=0 is an orthonormal basis of H (0, π) (see Exercise Problem 3.5).
Expand sin x in terms of this H 1 -basis, we obtain that
∞ n
2 X 4 2 X 4
sin x = − 2
cos 2kx = − lim cos 2kx ,
π π(4k − 1) π n→∞ π(4k 2 − 1)
k=1 k=1
3.5. 1-D HEAT EQUATIONS WITH PERIODIC BOUNDARY CONDITION141
where the limit is taken in the H 1 -topology, or equivalently,

n
2 X 4
lim sin x − + cos 2kx = 0.

π π(4k 2 − 1)
1
n→∞ H (0,π)
k=1
∂wk
Note that wk has the property that the derivative of wk , , vanishes at
∂x
the boundary points x = 0 and x = π for all k, but the derivative of sin x at
the boundary points does not vanish.
3.5 1-D heat equations with periodic boundary

condition
In this section, we consider the heat equation:
ut (x, t) − uxx (x, t) = f (x, t) in (0, 2π) × (0, T ) , (3.16a)

u(0, t) = u(2π, t) for all t ∈ (0, T ), (3.16b)
u(x, 0) = g(x) on (0, 2π) . (3.16c)
Condition (3.16b) is called the periodic boundary condition, which enables

us to treat solutions u(·, t) as a periodic function defined on R for all t ∈
[0, T ].
Z T Z Z T
2
We assume that |f (x, t)| dxdt = kf (·, t)k2L2 (T) dt < ∞, or f ∈
0 T 0
L2 (0, T ; L2 (T)), and g ∈ H 1 (T).
3.5.1 Formal approach
Assume that for all t ∈ [0, T ], u(·, t) ∈ L2 (T). Therefore, if dn (t) is the
Fourier coefficient of u(·, t), we can express u(x, t) as
∞ ∞
1 X X
u(x, t) = √ dk (t)eikx = dk (t)ek (x) .
2π k=−∞ k=−∞
Because of (3.16c), we must have dk (0) = ĝ(k). Moreover, for almost all
t ∈ [0, T ], f (·, t) ∈ L2 (T). Therefore,
∞
1 X ˆ
f (x, t) = √ fk (t)eikx for almost all t ∈ (0, T ) ,
2π k=−∞
where fˆk (t) is the Fourier coefficients defined by

1
Z
ˆ
fk (t) = √ f (x, t)e−ikx dx .
2π T
Suppose that we can switch the order of the differentiation and the summa-
tion, then
∞ ∞
1 X 0 1 X 2
ut (x, t) = √ dk (t)eikx , uxx (x, t) = − √ k dk (t)eikx ;
2π k=−∞ 2π k=−∞
thus by (3.16a), for almost all t ∈ [0, T ],

∞
1 X h 0 i
√ dk (t) + k 2 dk (t) − fˆk (t) eikx = 0 . (3.17)
2π k=−∞
Since {ek }∞
k=−∞ is maximal, we find that dk (t) solves the ODE
dk0 (t) + k 2 dk (t) = fˆk (t) . (3.18)
Together with the initial condition dk (0) = ĝ(k), we find that

Z t
−k2 t 2
dk (t) = e ĝ(k) + fˆk (s)e−k (t−s) ds
0
which implies that a solution u(x, t) can be written as

∞ Z t
1 X h −k2 t 2
i
u(x, t) = √ e ĝ(k) + fˆk (s)e−k (t−s) ds eikx . (3.19)
2π k=−∞ 0
3.5.2 Rigorous approach
Before proceeding, we state a very important theorem in the study of dif-

ferential equations.
Theorem 3.37 (The Gronwall inequality). Let x(t) be a non-negative, con-

tinuous function on the interval [0, T ]. If x(t) satisfies x 0 (t) ≤ M + Cx(t)
for all t ∈ [0, T ], then
M Ct
x(t) ≤ eCt x(0) + (e − 1) ∀ t ∈ [0, T ] . (3.20)
C
Proof. Multiplying both sides of the differential inequality by the integrating
factor e−Ct , we find that
d h −Ct i
e x(t) ≤ M e−Ct .
dt
§3.5 1-D heat equations with periodic boundary condition 143
The desired inequality is then obtained by integrating the inequality above

in time from 0 to t for some t ∈ [0, T ], and the detail is left to the readers.
Corollary 3.38. Let y(t) be a non-negative, integrable function on the in-

terval [0, T ]. If y(t) satisfies
Z t
y(t) ≤ M + C y(s)ds ∀ t ∈ [0, T ] , (3.21)
0
then
y(t) ≤ M eCt ∀ t ∈ [0, T ] .

Z t
Proof. Let x(t) = y(s)ds and apply Theorem 3.37 .
0
We truncate the Fourier presentation of g and f and find the solution to

the new problem.
n n
1 X ˆ 1 X
Let fn (x, t) = √ fk (t)eikx and gn (x) = √ ĝ(k)eikx . We
2π k=−n 2π k=−n
look for a solution un (x, t) to
unt (x, t) − unxx (x, t) = fn (x, t) in (0, 2π) × (0, T ) , (3.22a)

un (0, t) = un (2π, t) for all t ∈ (0, T ), (3.22b)
un (x, 0) = gn (x) on (0, 2π) . (3.22c)
The same procedure as the formal approach implies that

n Z t
1 X h −k2 t 2
i
un (x, t) = √ e ĝ(k) + fˆk (s)e−k (t−s) ds eikx
2π k=−n 0
is a solution to (3.22).
Energy estimates
Multiplying (3.22a) by un (x, t) and integrating over T,

Z Z Z
unt (x, t)un (x, t)dx − unxx (x, t)un (x, t)dx = fn (x, t)un (x, t)dx .
T T T
(3.23)
Integrating by parts in x implies that

Z
unxx (x, t)un (x, t)dx = −kunx (·, t)k2L2 (T) .
T
1∂
Since unt (x, t)un (x, t) = |un (x, t)|2 , integrating (3.23) over time interval
2 ∂t
(0, t), we obtain
t
1
Z
kun (·, t)k2L2 (T) + kunx (·, t)k2L2 (T)
2 0
Z tZ
1
= kun (·, 0)k2L2 (T) + fn (x, s)un (x, s)dxds .
2 0 T
By Hölder’s and Young’s inequality,

Z
fn (x, s)un (x, s)dx ≤ kfn (·, s)kL2 (T) kun (·, s)kL2 (T)
T
1 1
≤ kfn (·, s)k2L2 (T) + kun (·, s)k2L2 (T) .
2 2
As a consequence,
Z t
kun (·, t)k2L2 (T) + 2 kunx (·, t)k2L2 (T)
0
h Z t i Z t
2 2
≤ kgn kL2 (T) + kfn (·, s)kL2 (T) ds + kun (·, s)k2L2 (T) ds
0 0
h Z t i Z t
≤ kgk2L2 (T) + kf (·, s)k2L2 (T) ds + kun (·, s)k2L2 (T) ds ,
0 0
where the last inequality follows from the Bessel inequality. Therefore, by
Corollary 3.38 we find that
h Z t i
kun (·, t)k2L2 (T) ≤ kgk2L2 (T) + kf (·, s)k2L2 (T) ds et ,
0
which in turn leads to that

Z t h Z T i
2 2 2
kun (·, t)kL2 (T) + 2 kunx (·, s)kL2 (T) ds ≤ C kgkL2 (T) + kf (·, s)k2L2 (T) ds
0 0
for some constant C > 0 (which depends on T ). In terms of the Fourier

coefficients,
n h
X Z t i
2
|dk (t)| + 2 |k|2 |dk (s)|2 ds ≤ M < ∞ . (3.24)
k=−n 0
Since the right-hand side of (3.24) is independent of n, we conclude that

∞ h
X Z t i
|dk (t)|2 + 2 |k|2 |dk (s)|2 ds ≤ M < ∞ (3.25)
k=−∞ 0
which is the same as saying that

Z t
2
ku(·, t)kL2 (T) + 2 kux (·, s)k2L2 (T) ds ≤ M < ∞ ,
0
∞
1 X
where u(x, t) = √ dk (t)eikx is the limit of un (x, t).
2π k=−∞
Remark 3.39. Here we must clarify in what sense un converges to u. Since

X
kun (·, t) − u(·, t)k2H 1 (T) = (1 + |k|2 )|dk (t)|2 ,
|k|>n
by (3.24) we find that

Z T
lim kun (·, t) − u(·, t)k2H 1 (T) dt = 0 .
n→∞ 0
It is in this sense un converges to u.
In order to see u(x, t) is indeed a solution to (3.16), we prove that the

limit (under a proper sense) of (3.22) as n → ∞ is (3.16). Since fn → f in
L2 (0, T ; L2 (Ω)) and gn → g in L2 (T), it suffices to show that unt → ut and
unxx → uxx under certain sense. This amounts to show that
∞ ∞
∂ X X
dk (t)eikx = dk0 (t)eikx , (3.26a)
∂t
k=−∞ k=−∞
∞ ∞
∂2 X X
dk (t)eikx = − k 2 dk (t)eikx ; (3.26b)
∂x2
k=−∞ k=−∞
∂ ∂2
or in other words, and commute with the infinite sum.
∂t ∂x2
As discussed in Section 3.4.2, (3.26b) holds in the sense of L2 (T) if
∂
denotes the weak differentiation in x. To be more precise, if u(x, t) =
∂x
∞
1 X
√ dk (t)eikx ∈ H 2 (T) for almost all t ∈ (0, T ) and uxx denotes the
2π k=−∞
second weak (partial) derivatives of u in x, then
∞
1 X 2
uxx (x, t) = − √ k dk (t)eikx ,
2π k=−∞
where the infinite sum is understood in the sense of L2 (T); that is,
n
X ∞
X
2 ikx
k dk (t)e → k 2 dk (t)eikx in L2 (T) .
k=−n k=−∞
To see if u(·, t) ∈ H 2 (T) for almost all t ∈ (0, T ), we multiply (3.22a) by

unxx and the integrate over T to obtain that
Z Z
unt (x, t)unxx (x, t)dx − kunxx k2L2 (T) = fn (x, t)unxx (x, t)dx . (3.27)
T T
Integrating by parts in x implies

1d
Z Z Z
unt (x, t)unxx (x, t)dx = − untx (x, t)unx (x, t)dx = − |unx (x, t)|2 dx ;
T T 2 dt T
thus further integrating (3.27) over the time interval (0, t) leads to
Z t
1 2
kunx (·, t)kL2 (T) + kunxx (·, s)k2L2 (T) ds
2 0
Z tZ
1 2
= kunx (0, ·)kL2 (T) − fn (x, s)unxx (x, s)dxds
2 0 T
Z tZ
1
= kgnx k2L2 (T) − fn (x, s)ukxx (x, s)dxds .
2 0 T
By Hölder’s and Young’s inequality,

Z
fn (x, s)unxx (x, s)dx ≤ kfn (·, s)kL2 (T) kunxx (·, s)kL2 (T)
T
1 1
≤ kfn (·, s)k2L2 (T) + kunxx (·, s)k2L2 (T) .
2 2
Therefore, by kgnx kL2 (T) ≤ kgkH 1 (T) and kfn (·, t)kL2 (T) ≤ kf (·, t)kL2 (T) ,
Z t Z T
2 2 2
kunx (·, t)kL2 (T) + kunxx (·, s)kL2 (T) ds ≤ kgkH 1 (T) + kf (·, s)k2L2 (T) ds .
0 0
In terms of Fourier coefficients,

X n h Z t i Z T
|k|2 |dk (t)|2 + |k|4 |dk (s)|2 ds ≤ kgk2H 1 (T) + kf (·, s)k2L2 (T) ds .
k=−n 0 0
Since the right-hand side of the estimate above is independent of n, passing

n → ∞ implies that
∞ h
X Z t i Z T
|k|2 |dk (t)|2 + |k|4 |dk (s)|2 ds ≤ kgk2H 1 (T) + kf (·, s)k2L2 (T) ds .
k=−∞ 0 0
(3.28)
This shows u(·, t) ∈ H 2 (T) for almost all t ∈ (0, T ); thus (3.26b) holds for
almost all t ∈ (0, T ) provided that g ∈ H 1 (T) and f ∈ L2 (0, T ; L2 (T)).
Finally, since unxx → uxx and fn → f both in L2 (0, T ; L2 (T)), by (3.22a)
∞
1 X 0
we also conclude that unt converges to √ dk (t)eikx in L2 (0, T ; L2 (T));
2π k=−∞
thus (3.26a) is guaranteed. To summarize, we have the following
Theorem 3.40. Suppose that f ∈ L2 (0, T ; L2 (T)), and g ∈ H 1 (T). Then

∞ Z t
1 X h −k2 t 2
i
u(x, t) = √ e ĝ(k) + fˆk (s)e−k (t−s) ds eikx (3.29)
2π k=−∞ 0
satisfies
ut (x, t) − uxx (x, t) = f (x, t) in (0, 2π) for almost all t ∈ (0, T ) ,
u(0, t) = u(2π, t) for all t ∈ (0, T ),
u(x, 0) = g(x) on (0, 2π) .
Remark 3.41. When performing the energy estimates, we do not really

make use of the exact form of dn (t) :
Z t
−k2 t 2
dk (t) = e ĝ(k) + fˆk (s)e−k (t−s) ds .
0
In other words, as long as the ODE (3.18) has a solution (which is usually
provided by the fundamental theorem of ODE), (3.25) and (3.28) can be
obtained without the knowledge of the analytic form of dk (t).
3.5.3 The special case f = 0
There are some good properties for the solution u to the heat equation (with
periodic boundary condition) when there is no external forcing. We study
these properties in this sub-section.
Maximum principle
Multiplying the heat equation ut − uxx = 0 by pu|u|p−2 and then integrating

over T, we find that
d
Z
u(·, t) p p + p(p − 1) u2x (x, t)up−2 (x, t)dx = 0 .

dt L (T)
T
Integrating in time over the time interval (0, t) then implies that

u(·, t) p ≤ kgkLp (T) .
L (T)
Passing p to ∞, we find that

u(·, t) ∞ ≤ kgkL∞ (T) ∀t > 0. (3.30)
L (T)
This inequality reads that the absolute value of the solution never exceeds
the absolute value of the initial state, and is called the maximum principle
for the heat equation (with periodic boundary condition).
3.5.4 Decay estimates
Under the assumption f = 0; that is, we are in the situation that there is
no heat source in the environment,Zwe expect that the
Z solution/temperature
1
will converges to the average ḡ ≡ − g(x)dx ≡ g(x)dx as t → ∞. We
T 2π T
would like to study the convergence rate of u − ḡ.
By (3.29) with f = 0 and the Parseval identity,
2
X X
ku(·, t) − ḡk2L2 (T) = e−2k t |ĝ(k)|2 ≤ e−2t |ĝ(k)|2 ≤ e−2t kgk2L2 (T)
k6=0 k6=0
which implies that
u(·, t) − ḡ 2 ≤ e−t kgkL2 (T)

L (T)
∀t > 0.
Usually we are more interested in the case of t 1. In such a case, we

may evaluate u − ḡ in L∞ (T) and obtain that
u(·, t) − ḡ ∞ ≤ √1 −k2 t ≤ 1 kgkL1 (T) 2 2

X X
e−k (t−1) e−k

L (T)
e ĝ(k)
2π k6=0 2π
n6=0
(t ≥ 1) 1 −(t−1) X 2
≤ e kgkL1 (T) e−k ≤ Ce−t kgkL1 (T) ,
2π
k6=0
1
where we use the fact that sup ĝ(k) ≤ √ kgkL1 (T) to conclude the in-
k∈Z 2π
equality. Moreover, suppose that g is smooth so that u is smooth, then
∂ù X 2
`
(x, t) = e−k t ĝ(k)(ik)` eikx ;
∂x
k6=0
3.6. 1-D HEAT EQUATIONS WITH DIRICHLET BOUNDARY CONDITION149
thus for all k ∈ N, similar argument implies that

∂ù X 2
(·, t) ∞ ≤ e−(t−1) kgkL1 (T) e−k |k|` ≤ C` e−t kgkL1 (T) ∀t ≥ 1.

`
∂x L (T)
k6=0
This proves the following
Theorem 3.42. Let u be the solution to the heat equation (3.16) with f = 0
and g ∈ L1 (T). Then the `-th partial derivatives of u − ḡ with respect to x
decays exponentially to zero in the uniform sense.
3.6 1-D heat equations with Dirichlet boundary

condition
In this section, we consider the following initial-boundary value problem for

the heat equation
ut (x, t) = uxx (x, t) in (0, L) × (0, ∞) , (3.31a)

u(0, t) = u(L, t) = 0 for all t , (3.31b)
u(x, 0) = f (x) ∀ x ∈ [0, L] . (3.31c)
Because of the boundary condition (3.31b), we use the orthonormal basis

nr 2 ∞
kπx o∞ X kπx
sin . Assume that u(x, t) = dk (t) sin . Then
L L k=1 L
k=1
L
π2 k2 2 kπx
Z
dk0 (t) + dk (t) = 0 ∀t > 0, dk (0) = f (x) sin dx .
L2 L 0 L
Therefore, by solving dk (t),

∞
π 2 k2 kπx
dk (0)e− t
X
u(x, t) = L2 sin .
L
k=1
If f ∈ L2 (0, L), u defined above is a solution to (3.31).
3.7 Exercises
Problem 3.1. Prove Lemma 3.15.

Problem 3.2. Let f be a 2π-periodic Lipchitz function. Show that for n ≥ 2,

1 + 2 log n
kf − Fn+1 ∗ f kL∞ (T) ≤ kf kC 0,1 (T) (3.32)
2n
and
2
f − sn (f, ·) ∞ ≤ 2π(1 + log n) kf kC 0,1 (T) .

L (T)
(3.33)
n
Hint: For (3.32), apply the estimate
nn + 1 π o
Fn (x) ≤ min ,
2π 2(n + 1)x2
in the following inequality:
h
Z δ Z −δ Z π i
f (x) − Fn+1 ∗ f (x) ≤ + + f (x + y) − f (x)Fn+1 (y)dy
−δ −π δ
π
with δ = . For (3.33), use (3.9) and note that
n+1
inf kf − pkL∞ (T) ≤ kf − Fn ∗ f kL∞ (T) .
p∈Pn (T)
Problem 3.3. A function f : T → R is said to be piecewise C 1 if there

are finitely many disjoint open intervals Ii so that f ∈ C 1 (Ii ) for all i and
S¯
Ii = T. Show that Dn ∗ f converges to f uniformly as n → ∞ on any
i
compact subset of Ii .
Problem 3.4. In this problem, we are concerned with the following
Theorem 3.43 (Bernstein). Suppose that f is a 2π-periodic function such

that for some constant C and α ∈ (0, 1),
inf kf − pkL∞ (T) ≤ Cn−α

p∈Pn (T)
for all n ∈ N. Then f ∈ C 0,α (T).
Complete the following to prove the theorem.
1. Suppose that there is p ∈ Pn (T) such that
kp 0 kL∞ (T) > n , kpkL∞ (T) < 1 , and p 0 (0) = kp 0 kL∞ (T) .
π π
Choose γ ∈ − , such that sin(nγ) = −p(0) and cos(nγ) > 0, and
n n
π 1
define αk = γ + k+ for −n ≤ k ≤ n. Show that the function
n 2
r(x) = sin n(x − γ) − p(x) has at least one zeros in each interval
(αk , αk+1 ).
§3.7 Exercise 151
2. Let s ∈ N be such that such that 0 ∈ (αs , αs+1 ). Show that r has
at least 3 distinct zeros in (αs , αs+1 ) by noting that r0 (0) < 0 and
r(0) = 0.
3. Combining 1 and 2, show that
kp 0 kL∞ (T) ≤ nkpkL∞ (T) ∀ p ∈ Pn (T). (3.34)
4. Choose pn ∈ Pn (T) such that kf − pn k ≤ 2Cn−α for n ∈ N. Define

∞
P
q0 = p1 , and qn = p2n − p2n−1 for n ∈ N. Show that qn = f and
n=0
the convergence is uniform.
5. Show that kqn kL∞ (T) ≤ 6C2−nα . As a consequence, show that
qn (x) − qn (y) ≤ 6Cn2n(1−α) |x − y| and qn (x) − qn (y) ≤ 12C2−nα .

6. For any x, y ∈ T with |x − y| ≤ 1, choose m ∈ N such that 2−m ≤

|x − y| ≤ 21−m . Then use the inequality
∞
m−1
X X
f (x) − f (y) ≤ qn (x) − qn (y) + qn (x) − qn (y)
n=0 n=m

to show that f (x) − f (y) ≤ B|x − y|α for some constant B > 0.
Problem 3.5. Show that {wk }∞

k=0 defined in Remark 3.36 is an orthonormal
basis of H 1 (0, π).
Hint: Use the Parseval identity to show that {wk }∞
k=0 is a maximal or-
thonormal set of H 1 (0, π); that is, show that for all f ∈ H 1 (0, π),
Z π ∞
(f, wk )H 1 (0,π) 2 .
X
kf k2H 1 (0,π) = |f (x)|2 + |f 0 (x)|2 dx =

0 k=0
nr 2 o∞
You might need the fact that sin kx is an orthonormal basis of
π k=1
L2 (0, π).
Problem 3.6. Let f (x) = x on [−π, π]. Then f 0 (x) = 1 is certainly a

∞
L2 (−π, π)-function. However, you may want to check that |n|2 |fˆ(n)|2 =
P
n=−∞
∞, so by “definition”, it does not seem to be an H 1 (−π, π)-function. What
is wrong with the argument?
Problem 3.7. Show Remark 3.33.
Problem 3.8 (Generalized Gronwall inequality). Show that if a ∈ L1 (0, T )

is a non-negative function , and x(t) satisfies the following integral integral
inequality
Z t
x(t) ≤ M + a(s)x(s)ds .
0
Z t
Then x(t) ≤ M exp a(s)ds for all t ∈ [0, T ]. In particular, if x satisfies
0
x0 (t) ≤ b(t) + a(t)x(t)
for some a, b ∈ L1 (0, T ) and a ≥ 0, then

Z t
x(t) ≤ x(0) + kbkL1 (0,T ) exp a(s)ds ∀ t ∈ [0, T ] .
0
Problem 3.9. Use Fourier series to formally solve the following initial-
boundary value problem for the wave equation
utt (x, t) = c2 uxx (x, t) in (0, 1) × R ,

u(0, t) = u(1, t) = 0 for all t ,
u(x, 0) = f (x) , ut (x, 0) = g(x) ∀ x ∈ [0, 1] .
Derive the following two conservation laws from your Fourier series solution
and directly from the PDE:
d 1h
Z i
|ut (x, t)|2 + c2 |ux (x, t)|2 dx = 0 .
dt 0
Problem 3.10. Use Fourier series to formally solve the following initial-
boundary value problem for the Schrödinger equation
iut (x, t) = −uxx (x, t) in (0, 1) × R ,

u(0, t) = u(1, t) = 0 for all t ,
u(x, 0) = f (x) ∀ x ∈ [0, 1] .
Derive the following two conservation laws from your Fourier series solution
and directly from the PDE:
d 1 d 1
Z Z
|u(x, t)|2 dx = 0 , |ux (x, t)|2 dx = 0 .
dt 0 dt 0
§3.7 Exercise 153
Problem 3.11. Try using the Fourier series to solve
ut (x, t) = uxx (x, t) in (0, π) × (0, ∞) ,

u(0, t) = ux (π, t) = 0 for all t ,
u(x, 0) = f (x) ∀ x ∈ [0, π] .
The most important task is to look for a suitable basis that fits the boundary
condition.
Problem 3.12. Let (r, θ) be the polar coordinate on R2 .
(1) Show that a harmonic function u on Ω ⊂ R2 satisfies

1 1
(rur )r + 2 uθθ = 0 r > 0.
r r
(2) For α > 0, let Ωα be the wedge given in polar coordinates (r, θ) by
Ωα = {(r, θ) | 0 < r < 1 , 0 < θ < α} .
Based on the fact that the general solution to
r2 R 00 (r) + rR 0 (r) − s2 R(r) = 0
is of the form R(r) = C1 rs + C2 r−s , use the Fourier series to find a

bounded solution to the following boundary value problem
∆u = 0 in Ωα ,
u=0 on {θ = 0, α} ,
πθ
u = sin on {r = 1} .
α
u=0 u = sin πθ
α
∆u = 0
u=0
P
Hint: Suppose that u(r, θ) = k Rk (r)ek (θ), where {ek } forms an
orthonormal basis of L2 (0, α) satisfying certain boundary conditions
(you have to figure out what these boundary conditions are). Solve Rk
by finding an ODE for Rk .
(3) Find all α > 0 so that u ∈ C 2 (Ωα ).
Problem 3.13. Complete the following.
(1) Suppose that {en }∞ 2 ∞

n=1 is an orthonormal basis of L (0, `1 ) and {ẽm }m=1
is an orthonormal basis of L2 (0, `2 ). Show that {en (x)ẽm (y)}∞
n,m=1
forms an orthonormal basis of L2 ([0, `1 ] × [0, `2 ]).
Hint: Check the orthonormality and the maximality. For the maxi-
mality, check the Parseval identity.
(2) Solve the following PDE:
ut (x, y, t) − ∆u(x, y, t) = 0 (x, y) ∈ (0, π) × (0, π) , t > 0 ,

u(x, y, 0) = x(π − x) sin y (x, y) ∈ (0, π) × (0, π) ,
ux (0, y, t) = ux (π, y, t) = 0 y ∈ (0, π) , t > 0 ,
u(x, 0, t) = u(x, π, t) = 0 x ∈ (0, π) , t > 0 .
u=0
ut − ∆u = 0 ux = 0
ux = 0
u|t=0 = x(π − x) sin y
u=0
(3) Show that for all t ≥ 0, u from (b) satisfies
π π Z tZ π π
π6
Z Z Z
|u(x, y, t)|2 dxdy + 2 |Du(x, y, s)|2 dxdyds = .
0 0 0 0 0 60
Chapter 4
The Laplace and Poisson

Equations
Definition 4.1. On regions Ω ⊂ Rn , the Laplace operator ∆, also called the

Laplacian, is defined as
n
X ∂2
∆= .
i=1
∂x2i
Definition 4.2. A C 2 -function u satisfying ∆u = 0 is called a harmonic

function.
4.1 The fundamental solution
When Ω = Rn , the Laplace operator has radial symmetry, and we may

search for harmonic functions on Rn which depend only upon the radial
component. Letting
r = |x| ,
we look for a harmonic function u satisfying
u(x) = v(r) .
∂r xi
Since = , by the chain rule,
∂xi r
∂u dv ∂r xi ∂2u ∂ h 0 xi i x2i h 1 x2 i
= = v 0 (r) , = v (r) = v 00
(r) + v 0
(r) − 3i .
∂xi dr ∂xi r ∂x2i ∂xi r r2 r r
155
156 CHAPTER 4. The Laplace/Poisson Equations
Therefore,
n
X ∂2u n−1 0
∆u = = v 00 (r) + v (r) .
i=1
∂x2i r
Hence ∆u = 0 if and only if
n−1 0
v 00 + v = 0.
r
If we consider solutions away from r = 0 and suppose that v 0 (r) 6= 0, then
n−1
[log v 0 (r)]0 = .
r
This is a simple ordinary differential equation which we can directly integrate

to find that for r > 0,

 b log r + c (n = 2)
v(r) = b

n−2
+c (n ≥ 3) ,
r
where b and c are constants. These radially symmetric functions, harmonic
away from the origin, provide us with the singular integral kernels on Rn and
explicit representations for the solutions to the Poisson equation −∆u = f ,
at least when the forcing function f is “nice” enough.
Definition 4.3. The function

 1
 −
log |x| (n = 2)
2π

Φ(x) = 1

 (n ≥ 3) ,
(n − 2)ωn−1 |x|n−2
defined for x ∈ Rn , x 6= 0 , is the fundamental solution of Laplace’s equation,

where ωn−1 denotes the surface area of the unit ball in Rn and is defined by
2π n/2
ωn−1 = ,
Γ(n/2)
where Γ denotes the Gamma function.
Notation. We will use that notation F,i to denote ∂F/∂xi , while F,ij
denotes ∂ 2 F/∂xi ∂xj and similarly for higher-order partial derivatives.
4.1. THE FUNDAMENTAL SOLUTION 157
A direct computation shows that

−xi
Φ,i (x) = |x|−n ,
ωn−1
1 h i
Φ,ij (x) = − |x| δij + nxi xj |x|−n−2 ,
2
ωn−1
and we have the following derivative estimates:
1
|Φ,i (x)| ≤ |x|1−n , (4.1a)
ωn−1
n
|Φ,ij (x)| ≤ |x|−n , (4.1b)
ωn−1
|Dα Φ(x)| ≤ C(n, |α|)|x|2−n−|α| , (4.1c)
where Dα and |α| are the multi-index notation defined by the following
Definition 4.4 (Multi-index). An n-dimensional multi-index is a vector

α = (α1 , · · · , αn ) of non-negative integers. |α| is defined as the sum of αk
and α! is defined as the product of αk ! , i.e.,
n
X n
Y
|α| = αk and α! = αk ! .
k=1 k=1
The differential operator Dxα is defined by

∂ α1 ∂ αn
Dxα = α1 · · · .
∂x1 ∂xαnn
When the (spatial) variable is specified, we simply use Dα to denote Dxα .
4.1.1 Uniform and Hölder Continuous Functions
For Ω ⊂ Rn open, a function u : Ω → R is Lipschitz continuous if
|u(x) − u(y)| ≤ C|x − y| ∀ x, y ∈ Ω , (4.2)
where C is a constant that depends on Ω but not on the function u itself.

The inequality (4.2) provides a uniform modulus of continuity. The standard
example of functions which are Lipschitz continuous but not differentiable is
given by u(x) = |x|. It is interesting to refine this functional framework to
be able to discern the regularity of functions u(x) = |x|α for positive α ≤ 1.
We wish to understand how “cuspy” the graph of u is near the origin, for
example. To do so, we replace the difference quotient bound in (4.2) with

the following inequality:
|u(x) − u(y)| ≤ C|x − y|α ∀ x, y ∈ Ω . (4.3)
Functions which satisfy the inequality (4.3) are termed Hölder continuous
with exponent α.
Definition 4.5 (Continuous functions and compact support). For Ω ⊂ Rn ,

we let C 0 (Ω) denote the continuous functions on Ω, and we denote by C00 (Ω)
those functions in C 0 (Ω) with compact support contained in Ω.
Definition 4.6 (Bounded continuous functions). For Ω ⊂ Rn we set
C 0 (Ω) := {u : Ω → R | u is bounded and continuous} ,
with norm kukC 0 (Ω) = max |u(x)|. For integers k ≥ 0, we let C k (Ω) denote
x∈Ω
the functions possessing partial derivatives to all orders up to k which are
k (Ω) to denote the functions in
bounded and continuous on Ω. We use Cloc
C k (B) for all bounded balls contained in Ω.
Definition 4.7 (Hölder continuous functions). For u ∈ C 0 (Ω) and 0 < α ≤

1, we set
u ∈ C 0,α (Ω) := u ∈ C 0 (Ω) kukC 0 (Ω) + [u]C 0,α (Ω) < ∞ ,

where
|u(x) − u(y)|
[u]C 0,α (Ω) = sup .
x,y∈Ω |x − y|α
x6=y
The norm is kukC 0,α (Ω) = kukC 0 (Ω) + [u]C 0,α (Ω) .
Theorem 4.8. The space u ∈ C 0,α (Ω) endowed with the norm k · kC 0,α (Ω)
is a Banach space.
We leave the proof as an exercise for the reader.

We will denote C00,α (Ω) = C 0,α (Ω) ∩ C00 (Ω).
4.1.2 The Poisson equation −∆u = f in Rn
Our objective, here, is to produce explicit solutions to the Poisson equation

−∆u = f in Rn . We will show that convolution between the fundamental
solution Φ and the “forcing function” f is a solution to this problem.
Definition 4.9. We set

Z
u(x) = Φ(x − y)f (y)dy (4.4)
Rn
whenever Φ(x − ·)f (·) ∈ L1 (Rn ).
Lemma 4.10. Suppose that f is bounded and integrable with compact sup-
port. Then if u is given by (4.4), u ∈ C 1 (Rn ) and for any x ∈ Rn and
i = 1, ..., n,
Z
u,i (x) = Φ,i (x − y)f (y)dy .
Rn
Proof. Since f is bounded with compact support, the integral
Z
Ii (x) ≡ Φ,i (x − y)f (y)dy
Rn
is well-defined and continuous
Z (in x). It suffices to show that it is the
derivative of u(x) ≡ Φ(x − y)f (y)dy .
Rn
Let ρ : (0, ∞) → R be a smooth, monotone increasing function such that
(
1 if z ∈ (2, ∞) ,
ρ(z) = and |ρ0 | ≤ 2 ,
0 if z ∈ (0, 1) ,
and define
Z |x − y|
u (x) = ρ Φ(x − y)f (y)dy .
Rn
Note that since ρ is uniformly bounded and Φ ∈ L1 (Rn ) , by the dominated
convergence theorem, u → u uniformly as → 0 on compact subsets.

Furthermore, as Dx [ρ |x−y|
Φ(x − y)]f (y) is also integrable for > 0, we
may differentiate under the integral to find that
∂ h |x − y|
Z i
u ,i (x) = ρ Φ(x − y) f (y)dy
n ∂xi
ZR
xi − yi 0 |x − y|

= ρ Φ(x − y)f (y)dy
Rn |x − y|
Z |x − y|
+ ρ Φ,i (x − y)f (y)dy .
Rn
Therefore,
Z x − y |x − y|
i i 0
|u ,i (x) − Ii (x)| ≤ ρ Φ(x − y)f (y)dy

Rn |x − y|

Z |x − y|
+ 1 − ρ Φ,i (x − y)f (y)dy .

Rn
Note that |ρ| ≤ 1 and |ρ0 | ≤ 2 . Moreover, since spt(1 − ρ) ⊆ [0, 2] , it follows
that spt 1 − ρ · ⊆ [0, 2] . Similarly, since spt(ρ0 ) ⊆ [1, 2] , we see that

spt(ρ0 · ⊆ [, 2] . As a consequence, for the case n ≥ 3 (the case n = 2 is

left as an exercise for the reader) ,
|u ,i (x) − Ii (x)|

2 kf kL∞ (Ω) kf kL∞ (Ω)
Z Z
−n+2
≤ |x − y| dy + |x − y|1−n dy
<|x−y|<2 ωn−1 |x−y|<2 (n − 2)ωn−1
Z 2 Z 2
kf kL∞ (Ω) h 2
Z Z i
≤ rdrdSz + drdSz
ωn−1 |z|=1 |z|=1 0
h 2 Z 2 Z 2 i
= kf kL∞ (Ω) rdr + dr → 0 as → 0 ;
0
hence u ,i → Ii uniformly as → 0 .
Finally, by the uniform convergence of u to u and u ,i to Ii as → 0 ,
we conclude that
x x
∂u
h Z i Z
u(x) = lim u (x) = lim u (x0 ) + dxi = u(x0 ) + Ii dxi ;
→0 →0 x0 ∂xi x0
thus u,i = Ii .
Remark 4.11. Given that DΦ is integrable near the origin, it is possible

to compute the first partial derivatives of u by taking the limit of a sequence
of difference quotients of u. On the other hand, since D2 Φ is not integrable
near the origin, analysis of second partial derivatives of u require some sort
of limiting process, wherein the singular behavior at |x| = 0 is either excised
or regularized.
For example, we might consider removing a small ball near the origin,
and defining an approximation to u as follows:
Z
ũ (x) ≡ Φ(x − y)f (y)dy ,
Rn \B(x,)
which makes ũ a differentiable function. However, as the domain of inte-

gration also depends upon x, differentiation of ũ becomes a bit complicated,
requiring a change of variables. To avoid this procedure, one alternative is
the introduction of the cut-off function ρ (introduced in the above proof ),
which has the similar affect of removing the singular region, without any
difficulties in differentiation.
Theorem 4.12. Suppose that Ω ⊂ Rn , f ∈ C00,α (Ω) with 0 < α ≤ 1 , and

suppose that u is given by (4.4). Then u ∈ C 2 (Rn ) , and for any x ∈ Rn ,
and i, j = 1, ..., n,
Z Z

u,ij (x) = Φ,ij (x − y) f (y) − f (x) dy − f (x) Φ,i (x − y)Nj dSy , (4.5)
Ω0 ∂Ω0
where Ω0 is any bounded, smooth domain containing Ω . In particular,
−∆u = f in Rn . (4.6)
Remark 4.13. Before starting the proof of Lemma 4.12, we explain why
a formula like (4.5) is well-defined. Note that as Dα Φ is not integrable
for |α| ≥ 2, ZΦ,ij (x − ·)f (·) 6∈ L1 (Rn ) even if f ∈ C0∞ (Rn ). Nevertheless,
Φ,ij (x − y) f (y) − f (x) dy is well-defined for all x ∈ Rn

the integral
Rn
due to the Hölder continuity of f . In particular, it is essential that second-
derivatives of Φ are multiplying the difference [f (y) − f (x)] – the Hölder
continuity of f “cancels” the singular nature of D2 Φ near the origin, at
least enough so that the integral converges.
The presence of the boundary integral on the right-hand side of (4.5) is
necessary in order to cancel the effect of the subtraction of f (x) from f (y).
Proof of Theorem 4.12. To see that (4.6) follows from (4.5), notice that
∆u = u,ii and that according to (4.5),
Z Z

u,ii (x) = Φ,ii (x − y) f (y) − f (x) dy − f (x) Φ,i (x − y)Ni dSy .
Ω0 ∂Ω0
Since Φ,ii (x − y) = 0 if x 6= y , and Φ,ii (x − ·) f (·) − f (x) ∈ L1 (Rn ) , the

first integral on the right-hand side vanishes. Thus

Z
u,ii (x) = −f (x) Φ,i (x − y)Ni dSy . (4.7)
∂Ω0
Choose R > 0 sufficiently large (for example, R = diam(Ω) + 1) so that

Ω0 ⊂⊂B(x, R), and let A = B(x, R) − Ω0 . Since ∆Φ(x − ·) = 0 in A, the
divergence theorem implies that
Z Z Z
0= ∆Φ(x−y)dx = DΦ(x−y)·n(y)dSy + DΦ(x−y)·n(y)dSy ,
A ∂B(x,R) ∂Ω0
where n denotes the outward-pointing unit normal to ∂A . Since n = −N

on ∂Ω0 , substitution into (4.7) shows that
Z
u,ii (x) = −f (x) Φ,i (x − y)Ni dSy
∂B(x,R)
−(xi − yi ) yi − x i
Z
= −f (x) |x − y|−n dSy
∂B(x,R) ωn−1 R
1
Z
= −f (x) R1−n dSy = −f (x) .
ωn−1 ∂B(x,R)
Next, we establish (4.5). Following the proof of Lemma 4.10, we define

Z |x − y|
i
v (x) = ρ Φ,i (x − y)f (y)dy .
Rn
The derivative of this integrand has an L1 (Rn ) dominating function, so the
dominated convergence theorem allows to differentiate under the integral.
We thus find that
∂ |x − y| i |x − y|
Z h Z
vi ,j (x) = ρ Φ,i (x − y)f (y)dy + ρ Φ,ij (x − y)f (y)dy
Ω0 ∂xj Ω0
∂ |x − y| i
Z h

= ρ Φ,i (x − y) f (y) − f (x) dy
Ω0 ∂xj
∂ |x − y| i
Z h
− f (x) ρ Φ,i (x − y)dy
Ω0 ∂yj
Z |x − y|
+ ρ Φ,ij (x − y) f (y) − f (x) dy
Ω0
Z |x − y|
+ f (x) ρ Φ,ij (x − y)dy .
Ω0
We will show that vi ,j converges to the right-hand side of (4.5) uniformly.
∂
Since Φ,ij (x − y) = − Φ,i (x − y) , integration by parts shows that
∂yj
Z |x − y|
ρ Φ,ij (x − y)dy
Ω0
∂ |x − y| i |x − y|
Z h Z
= ρ Φ,i (x − y)dy − ρ Φ,i (x − y)Nj dSy ,
Ω0 ∂yj ∂Ω0
4.2. GREEN’S FUNCTION 163
and hence that

Z h ∂ |x − y| i
vi ,j (x) =

ρ Φ,i (x − y) f (y) − f (x) dy
Ω0 ∂xj
Z |x − y|
+ ρ Φ,ij (x − y) f (y) − f (x) dy
Ω0
Z |x − y|
− f (x) ρ Φ,i (x − y)Nj dSy .
∂Ω0
Following the proof of Lemma 4.10, by the Hölder continuity of f ,
Z h ∂ |x − y| i
ρ Φ,i (x − y) f (y) − f (x) dy

Ω0 ∂xj

Z ∂ |x − y|
≤ ρ Φ,i (x − y)f (y) − f (x)dy

B(x,R) ∂xj

Z x − y |x − y|
i i 0
≤ ρ Φ,i (x − y)f (y) − f (x)dy

<|x−y|<2 |x − y|

Z 2[f ]C 0,α (Ω)
≤ |x − y|1−n+α dy
<|x−y|<2 ω n−1
2[f ]C 0,α (Ω) Z Z 2
= rα drdSz → 0 as → 0 .
ωn−1 |z|=1
Here we note that since R can be chosen independent of x ∈ Ω , the conver-
gence above is in fact uniform in x . Similarly, by (4.1b),
Z |x − y|
1 − ρ Φ,ij (x − y) f (y) − f (x) dy

Ω0
n[f ]C 0,α (Ω) Z Z 2
≤ rα−1 drdSz → 0 as → 0 ;
ωn−1 |z|=1 0
and again the convergence above is uniform in x ∈ Ω . Consequently, vi ,j

converges to the right-hand side of (4.5) uniformly as → 0 .
It remains to show that vi ,j → u,ij as → 0. From Lemma 4.10, vi → u,i
uniformly as → 0, so using the fundamental theorem of calculus (as in the
proof of Lemma 4.10), we indeed see that u,ij must be equal to the right-
hand side of (4.5).
4.2 Green’s function
For a point y ∈ Ω, the function Φ(x − y) is harmonic if x 6= y . Therefore,

letting v(x) = Φ(x − y) in Green’s second identity (1.2),
∂u ∂Φ
Z Z Z
Φ(x − y)∆u(x)dx = Φ(x − y) (x)dSx − u(x) (x − y)dSx
Ωk ∂Ωk ∂N ∂Ωk ∂N
where Ωk = Ω\B(y, k1 ) . For k big enough, ∂Ωk = ∂Ω∪∂B(y, k1 ) . Therefore,

∂Φ ∂u ∂u
Z Z Z
u(x) (x − y)dSx = Φ(x − y) (x)dSx + Φ(x − y) (x)dSx
1
∂B(y, k ) ∂N ∂Ω ∂N 1
∂B(y, k ) ∂N
∂Φ
Z Z
− u(x) (x − y)dSx − Φ(x − y)∆u(x)dx ,
∂Ω ∂N 1
Ω\B(y, k )
here we have to note that the unit normal on ∂B(y, k1 ) points to the center
y. For x ∈ ∂B(y, k1 ) ,
1


 log k if n = 2
 2π
Φ(x − y) = ,
k n−2
if n ≥ 3



(n − 2)ωn−1
∂Φ k n−1
(x − y) = DΦ(x − y) · k(y − x) = .
∂N ωn−1
Therefore, as k → ∞ ,
∂Φ ∂u
Z Z
u(x) (x − y)dSx → u(y) and Φ(x − y) (x)dSx → 0 .
1
∂B(y, k ) ∂N 1
∂B(y, k ) ∂N
Z Z
Moreover, Φ(x − y)∆u(x)dx → Φ(x − y)∆u(x)dx as k → ∞ . As a con-
Ωk Ω
sequence, u ∈ C 2 (Ω) satisfies
∂u ∂Φ
Z Z Z
u(x) = Φ(y − x) (y)dSy − u(y) (y − x)dSy − Φ(y − x)∆u(y)dy .
∂Ω ∂N ∂Ω ∂N Ω
(4.8)
Z
Remark 4.14. The integral Φ(x − y)f (y)dy is called the Newtonian
Ω
potential with density f .
Given the formula (4.8) it is tempting to believe that the equation
−∆u = f in Ω,
u=g on ∂Ω ,
∂u
=h on ∂Ω
∂N
has a solution
∂Φ
Z Z Z
u(x) = Φ(x − y)h(y)dSy − g(y) (y − x)dSy + Φ(y − x)f (y)dy .
∂Ω ∂Ω ∂N Ω
This is, in fact, not the case, and we shall examine this in great detail below.
4.3. PROPERTIES OF HARMONIC FUNCTIONS 165
4.3 Properties of harmonic functions

4.3.1 Mean-value property
Now we consider harmonic functions on an open set Ω ⊆ Rn .
Theorem 4.15. If u ∈ C 2 (Ω) is harmonic, then

Z Z
u(x) = − u(y)dSy = − u(y)dy (4.9)
∂B(x,r) B(x,r)
for each ball B(x, r) ⊆ Ω .
Proof. We begin by proving the first equality in (4.9). By the divergence

theorem,
∂u ∂
Z Z
0= (y)dSy = rn−1 u(x + rw)dSw
∂B(x,r) ∂N ∂B(0,1) ∂r
∂ n−1 ∂
Z h Z i
n−1 1−n
=r u(x + rw)dSw = r r u(y)dSy .
∂r ∂B(0,1) ∂r ∂B(x,r)
Therefore,
Z Z
r1−n u(y)dSy = lim r1−n u(y)dSy = ωn−1 u(x) . (4.10)
∂B(x,r) r→0+ ∂B(x,r)
The second equality in (4.9) is obtained by integrating (6.44) in r .
Exercise: If u ∈ C 2 (Ω) satisfying ∆u ≥ 0 (or ∆u ≤ 0), show that

Z Z
u(x) ≤ − u(y)dSy or u(x) ≥ − u(y)dSy ,
∂B(x,r) ∂B(x,r)
Z Z
u(x) ≤ − u(y)dy or u(x) ≥ − u(y)dy .
B(x,r) B(x,r)
A function u ∈ C 2 (Ω) is called a sub-harmonic/super-harmonic function if

∆u ≥ 0/≤ 0.
Theorem 4.16 (Converse to mean-value property). If u ∈ C 2 (Ω) satisfies

(4.9) for each B(x, r) ⊆ Ω , then u is harmonic.
Proof. If ∆u 6= 0, there exists some ball B(x, r) ⊂ Ω on which ∆u > 0 (or

R
perhaps ∆u < 0) in B(x, r). But this would imply that 0 = −B(x,r) ∆u(y)dy >
0, a contradiction.
Theorem 4.17. If u ∈ L1loc (Ω) satisfies the mean-value property (4.9) for
each B(x, r) ⊆ Ω , then u ∈ C ∞ (Ω) .
Proof. Let η be the standard mollifier defined in Definition 2.25, and let
u = η ∗ u in Ω . Then
Z
1
Z |x − y|
u (x) = η (|x − y|)u(y)dy = n η u(y)dy
Ω B(x,)
1 r
Z Z
= n η( )u(x + rw)rn−1 dSw dr
0 ∂B(0,1)
Z
1 r
= n ωn−1 u(x) η( )rn−1 dr = u(x) .
0
Therefore, u = u in Ω .
Corollary 4.18. A harmonic function is a C ∞ -function.
Corollary 4.19. The limit of a uniformly convergent sequence of harmonic

functions is harmonic.
4.3.2 Maximum principles
Theorem 4.20 (Strong maximum principle). Suppose that u ∈ C 2 (Ω) ∩

C 0 (Ω) is harmonic within a bounded domain Ω .
(1) Then max u = max u .

Ω ∂Ω
(2) Furthermore, if Ω is connected and there exists a point x0 ∈ Ω such

that u(x0 ) = max u , then u = u(x0 ) within Ω .
Ω
Proof. Let M = max u (the boundedness of Ω implies that such an M

Ω
exists), and A = {x ∈ Ω | u(x) = M } . The continuity of u implies that A
is closed (relative to Ω). If A is empty, then the maximum of u is attained
on ∂Ω , so we may assume that A is not empty.
Let x ∈ A , and r > 0 be such that B(x, r) ⊆ Ω . According to (4.9)
Z
M = u(x) = − u(y)dy ≤ M .
B(x,r)
Equality can only hold if u(y) = u(x) for all y ∈ B(x, r); thus, B(x, r) ⊆ A,
and hence A is also open in Ω . It must be that A = Ω since it is the only
subset of Ω which is both open and closed in Ω .
§4.3 Properties of harmonic functions 167
Remark 4.21. In the statement of Theorem 4.20, it suffices to assume that

∆u ≥ 0.
Exercise: Let u be sub-harmonic/super-harmonic in Ω . Then
(1) Then max u = max u/ min u = min u .

Ω ∂Ω Ω ∂Ω
(2) If Ω is connected and there exists a point x0 ∈ Ω such that u(x0 ) =

max u/u(x0 ) = min u , then u = u(x0 ) within Ω .
Ω Ω
Corollary 4.22 (Uniqueness). Let g ∈ C 0 (∂Ω) , f ∈ C 0 (Ω) . Then there

exists at most one solution u ∈ C 2 (Ω)∩C 0 (Ω) to the boundary-value problem
−∆u = f in Ω,
u=g on ∂Ω .
Corollary 4.23 (Comparison). For u, v ∈ C 2 (Ω) ∩ C 0 (Ω), if ∆u ≥ 0,

∆v = 0 in Ω, and u = v on ∂Ω, then u ≤ v in Ω. (For this reason, we call
u subharmonic.
4.3.3 The Harnack inequality
Theorem 4.24. Let u be a non-negative harmonic function in Ω . Then for

any bounded sub-domain Ω0 ⊂⊂Ω , there exists a constant C depending only
on n , Ω0 and Ω such that
max
0
u ≤ C min
0
u.
Ω Ω
Proof. Let y ∈ Ω , B(y, 4R) ⊆ Ω . Then for any two points x1 , x2 ∈ B(y, R) ,
we have
1
Z Z
u(x1 ) = − u(x)dx ≤ u(x)dx ,
B(x1 ,R) |B(0, R)| B(x1 ,2R)
1
Z Z
u(x2 ) = − u(x)dx ≥ u(x)dx .
B(x2 ,3R) |B(0, 3R)| B(x1 ,2R)
Consequently, max u ≤ 3n min u . The general case follows from connect-

B(y,R) B(y,R)
m
ing any two points in Ω0 by an arc Γ so that Γ ⊆
S
B(xk , 4R) ⊆ Ω for
k=1
some xk ∈ Γ and R > 0 . The number m only depends on Ω0 and Ω , so the
constant C depends only on n , Ω0 and Ω .
4.3.4 Local estimates
Theorem 4.25. Assume u is harmonic in Ω and U ⊂⊂Ω . Then for any

multi-index α we have
n|α| |α|
sup |Dα u(x)| ≤ sup |u(x)| , (4.11)
x∈U d x∈Ω
where d = dist(U, ∂Ω) .
Proof. Suppose that α = ei = ( 0, · · · , 0 , 1, 0, · · · , 0) . Since Dα u is har-

| {z }
(i − 1) zeros
monic,
1 1
Z Z
α α
D u(x) = D u(y)dy = div(uei )dy
|B(x, r)| B(x,r) |B(x, r)| B(x,r)
1 |∂B(x, r)|
Z
= u(y)Ni (y)dSy ≤ max |u(y)| .
|B(x, r)| ∂B(x,r) |B(x, r)| y∈B(x,r)
[
As a consequence, with Ur denoting the set B(x, r) ,
x∈U
n
max |Dα u(x)| ≤ max |u(x)| .
x∈U r x∈Ur
The general result follows from applying the above inequality |α| times with
d
r= .
|α|
Remark 4.26. Inequality (4.11) is also called the gradient estimate for
harmonic functions.
4.3.5 Regularity of weakly harmonic functions
Theorem 4.27. For Ω ⊂ Rn , suppose that u ∈ L1loc (Ω) and satisfies

Z
u(x)∆φ(x)dx = 0 ∀ φ ∈ C02 (Ω) .
Ω
Then u is harmonic in Ω.
Proof. Without loss of generality, we may assume that u ∈ L1 (Ω).

Choose > 0 sufficiently small so that φ := η ∗ φ ∈ C0∞ (Ω), where η
denotes the standard mollifiers given in Definition 2.25. By assumption,
Z Z
0= u(x)∆φ (x)dx = u (x)∆φ(x)dx . (4.12)
Ω Ω
§4.3 Properties of harmonic functions 169
Since u is smooth, we can integrate by parts to find that

Z
∆u (x)φ(x)dx ∀ φ ∈ C02 (Ω) .
Ω
Thus, ∆u = 0 in Ω, so u is harmonic.

By Young’s inequality the sequence u is uniformly bounded in L1 (Ω):
ku kL1 (Ω) ≤ kη kL1 (Ω) kukL1 (Ω) = kukL1 (Ω) .
Since u is harmonic,
1
Z

u (x) = u (y)dy ,
ωn Rn B(x,R)
which implies that

1
|u (x)| ≤ ku kL1 (Ω) .
ωn Rn
The local gradient estimate (4.11) then provides the inequality
sup |Dα u(x)| ≤ C · sup |u | ≤ CkukL1 (Ω) .

x∈U Ω
We have therefore shown that Dα u is bounded and equicontinuous on

any U ⊂⊂Ω. By the Arzela-Ascoli Compactness Theorem 2.110, there exists
0
some subsequence 0 such that u → v ∈ C 2 (U ).
0
On the other hand, u → u in L1 (U ) so u = v on U . By pushing ∂U
closer to ∂Ω, we see that u is smooth in Ω. Hence for all φ ∈ C02 (Ω),
Z Z
0= u(x)∆φ(x)dx = ∆u(x)φ(x)dx ,
Ω Ω
so that u is harmonic.
4.3.6 Liouville’s Theorem
Theorem 4.28. Suppose u : Rn → R is harmonic and bounded. Then u is

constant.
Proof. By gradient estimate (4.11), the derivatives Dα u(x) have to vanish

for all x ∈ Rn and multi-index α . In particular, this implies that u is
constant along any lines, so u is constant.
Corollary 4.29. Let n ≥ 3 and f ∈ C 0,α (Rn ) with compact support in Rn .

Then any bounded solution of
−∆u = f in Rn
has the form

Z
u(x) = Φ(x − y)f (y)dy + C
Rn
for some constant C .
4.3.7 Analyticity
Theorem 4.30. Assume u is harmonic in Ω . Then u is analytic in Ω ,

that is, for each x0 ∈ Ω ,
X Dα u(x0 )
u(x) = (x − x0 )α
α
α!
within some ball B(x0 , r) .
4.4 Computing Green’s functions
For fixed x ∈ Ω , suppose that there is a harmonic function Φ̃x (y) such that
∆y Φ̃x (y) = 0 for all y ∈ Ω and
Φ̃x (y) = Φ(y − x) ∀ y ∈ ∂Ω .
Then by Green’s second identity,

∂u ∂ Φ̃x i
Z Z h
Φ̃x (y)∆u(y)dy = Φ̃x (y) (y) − u(y) (y) dSy . (4.13)
Ω ∂Ω ∂N ∂N
By (4.8) and (4.13), we obtain that if u ∈ C 2 (Ω) ∩ C 0 (Ω) , then
∂G
Z Z
u(x) = − u(y) (x − y)dSy − G(x, y)∆u(y)dy , (4.14)
∂Ω ∂N Ω
where G(x, y) = Φ(y −x)− Φ̃x (y). The function G(x, y) is called the Green’s
function for the domain Ω.
Theorem 4.31 (Symmetry of Green’s function). For all x , y ∈ Ω , x 6= y ,

we have
G(x, y) = G(y, x) .
§4.4 Computing Green’s functions 171
Proof. Fix x, y ∈ Ω , x 6= y . Define v(z) = G(x, z) and w(z) = G(y, z) . The

goal is to show that v(y) = w(x) .
Letting Ω = Ω−(B(x, )∪B(y, )), and applying Green’s second identity
(1.2), we obtain that
Z h ∂v ∂w i
Z h ∂w ∂v i
w− v dSz = v− w dS ,
∂B(x,) ∂N ∂N ∂B(y,) ∂N ∂N
where N denotes the inward-pointing unit normal on ∂B(x, ) ∪ ∂B(y, ) .

Passing → 0 , the left-hand side converges to w(x) while the right-hand
side converges to v(y) .
4.4.1 The case Ω = Rn−1 × R+
For x ∈ Rn+ ≡ Rn−1 × R+ , let x̃ = (x1 , · · · , xn−1 , −xn ) . Then Φ̃x (y) =
Φ(y − x̃) is harmonic in Rn+ , and Φ̃x (y) = Φ(y − x) for all y ∈ ∂Rn+ .
Therefore, the Green’s function is given by
G(x, y) = Φ(y − x) − Φ(y − x̃) ,
the outward unit normal to ∂Rn+ is N = −en = (0, ..., 1) so ∂G/∂N =

−∂G/∂yn , and
∂G −1 2xn
(x, y) = ,

∂N yn =0 ωn−1 |x − y|n
and Green’s representation formula (4.14) suggests that the (bounded) so-
lution to
∆u = 0 in Rn+ , (4.15a)
u=f on ∂Rn+ (4.15b)
is
∂G 2xn f (y)
Z Z
u(x) = − f (y) (x, y)dSy = dy (∀ x ∈ Rn+ ) . (4.16)
n
∂R+ ∂N ωn−1 Rn−1 |x − y|n
2xn
With K(x, y) = , the equation (4.16) can be written as
ωn−1 |x − y|n
Z
u(x) = K(x, y)f (y)dSy .
∂Rn
+
The function K is termed the Poisson kernel for Rn+ , and (4.16) is called
the Poisson integral formula.
It remains to verify that (4.16) provides a solution u ∈ C 2 (Rn+ ) ∩ C 0 (Rn+ )
with prescribed boundary condition.
Theorem 4.32. Assume that f ∈ C 0 (∂Rn+ ) ∩ L∞ (∂Rn+ ) , and u is given by

(4.16). Then u ∈ C 2 (Rn+ ) ∩ C 0 (Rn+ ) satisfies (4.15a) and (4.15b).
Proof. Since K(x, y) is smooth if x 6= y , (4.16) indeed shows that u is

smooth . In particular, u ∈ C 2 (Rn+ ) .
Next, we note that the Poisson kernel is normalized; namely,
2xn 1
Z Z
K(x, y)dSy = n
dy = 1 ∀ x ∈ Rn+ . (4.17)
n
∂R+ ωn−1 R n−1 |x − y|
The identity (4.17) is an immediate consequence of (4.14) with u ∈ C 2 (Rn+ )∩

C 0 (Rn+ ) taken to be u(x) = 1.
Let z ∈ ∂Rn+ . Since f ∈ C 0 (∂Rn+ ) , given > 0 , there exists δ > 0 such
that

|f (x) − f (z)| < whenever |x − z| < 2δ .
2
Then if |x − z| < δ , x ∈ Rn+ ,
Z
|u(x) − f (z)| = K(x, y) f (y) − f (z) dSy

∂Rn
+
Z Z
≤ K(x, y)|f (y) − f (z)|dSy + 2 max
n
|f | K(x, y)dSy
∂Rn ∂R+ ∂Rn
+ ∩B(z,2δ) + \B(z,2δ)

Z
≤ + 2 max |f | K(x, y)dSy .
2 ∂Rn
+ ∂Rn
+ \B(z,2δ)
| {z }
≡I
If y ∈ ∂Rn+ \B(z, 2δ) , |y − z| > 2δ . So if |x − z| < δ1 ≤ δ ,

1
|y − z| ≤ |y − x| + |x − z| ≤ |y − x| + |y − z|
2
which implies |y − x| ≥ 12 |y − z| . As a consequence,
2n+2 xn Cxn
Z
I≤ max |f | n
dSy ≤
ωn−1 ∂R+ n
∂Rn
+ \B(z,2δ)
|y − z| δ
for some constant C < ∞ . Choose δ1 even smaller so that 2Cδ1 < δ , then
|u(x) − f (z)| < whenever |x − z| < δ1 . This proves u ∈ C 0 (Ω) .
§4.4 Computing Green’s functions 173
4.4.2 The case Ω = B(0, 1) or Ω = B(0, R)

x
For x ∈ B(0, 1) , let x̃ = , and Φ̃x (y) = Φ(|x|(y − x̃)) . Then Φ̃x (y)
|x|2
is harmonic in Ω , i.e., ∆y Φ̃x (y) = 0 for all x, y ∈ Ω . Moreover, Φ̃x (y) =
Φ(y − x) for all y ∈ ∂B(0, 1) . Therefore, the Green’s function for the unit
ball is
G(x, y) = Φ(x − y) − Φ(|x|(y − x̃)) .
Using (4.14), we find that the solution to
∆u = 0 in B(0, 1) ,
u=f on ∂B(0, 1)
is
∂G 1 − |x|2 f (y)
Z Z
u(x) = − f (y) (x, y)dSy = dSy .
∂B(0,1) ∂N ωn−1 ∂B(0,1) |x − y|n
since
∂G −1 1 − |x|2
(x, y) = .

∂N y∈∂B(0,1) ωn−1 |x − y|n
By a change of variables, the solution to
∆u = 0 in B(0, R) , (4.18a)
u=f on ∂B(0, R) , (4.18b)
is then
R2 − |x|2 f (y)
Z
u(x) = dSy . (4.19)
ωn−1 R ∂B(0,R) |x − y|n
Similar to Theorem 4.32, we have that
Theorem 4.33. Assume that f ∈ C 0 (∂B(0, R)) , and u is given by (4.19).

Then u ∈ C 2 (B(0, R)) ∩ C 0 (B(0, R)) satisfies (4.18a) and (4.18b).
The function
R2 − |x|2 1
K(x, y) =
ωn−1 R |x − y|n
is the Poisson kernel for the ball B(0, R) .
4.5 Perron’s method and solutions to the Poisson

problem1
In this section, we prove the existence of solutions to
−∆u = f in Ω, (4.20a)
u=g on ∂Ω . (4.20b)
using Perron’s method under the assumption that f ∈ C 0,α (Ω) and g ∈
C 0 (∂Ω) .
First, we extend f to whole Rn with compact support so that the exten-
sion, still denoted by f , belongs to C 0,α (Rn ) . Let ϕ = Φ∗f , and v = u−ϕ .
By Lemma 4.12, −∆ϕ = f , v is harmonic. So, v solves
−∆v = 0 in Ω, (4.21a)
v =g−ϕ≡ψ on ∂Ω . (4.21b)
As long as we know how to solve the Dirichlet problem (4.21), we obtain a

solution to (4.20) by summing ϕ and the solution to (4.21). Therefore, we
concentrate on how (4.21) is solved.
First we generalize the notion of sub-harmonic function. Recall that a
function w ∈ C 2 (Ω) is sub-harmonic if ∆w ≥ 0 in Ω , and
Z
w(ξ) ≤ − w(y)dSy ≡ Mw (ξ, ρ) . (4.22)
∂B(ξ,ρ)
A function w ∈ C 0 (Ω) is called sub-harmonic, if for each ξ ∈ Ω , (4.22)

holds for all ρ > 0 such that B(ξ, ρ)⊂⊂Ω . Let σ(Ω) denote the space of all
sub-harmonic functions on Ω , and given ψ ∈ C 0 (∂Ω) , define
n o
σψ (Ω) = w ∈ σ(Ω) ∩ C 0 (Ω) w ≤ ψ on ∂Ω .

σψ (Ω) is non-empty since the constant function w ≡ c belongs to σψ (Ω) if

c ≤ inf ψ(x) .
x∈∂Ω
For u ∈ C 0 (Ω) and B(ξ, ρ)⊂⊂Ω , we define uξ,ρ , the harmonic lifting of
u, as the function in C 0 (Ω) for which
∆x uξ,ρ (x) = 0 in B(ξ, ρ) ,

uξ,ρ (x) = u(x) in Ω\B(ξ, ρ) .
1
The reader may skip this section on the first reading
§4.5 Perron’s method and solutions to the Poisson problem 175
Claim: For u ∈ σ(Ω) and B(ξ, ρ)⊂⊂Ω , u(x) ≤ uξ,ρ (x) for all x ∈ Ω , and
uξ,ρ ∈ σ(Ω) .
Proof. It suffices to show that
uξ,ρ (x) ≤ Muξ,ρ (x, r) ∀ r > 0 such that B(x, r)⊂⊂Ω . (4.23)
If B(x, r) ⊆ B(ξ, ρ) , since uξ,ρ is harmonic in B(ξ, ρ) , (4.23) holds because of

the mean-value property for the harmonic functions. If B(x, r)∩B(ξ, ρ) = ∅ ,
then uξ,ρ = u , so (4.23) holds because of (4.22). Other than these two
cases, let w be the harmonic function satisfying w = uξ,ρ on ∂B(x, r) . On
∂B(x, r) ∩ B(ξ, ρ) , w = uξ,ρ ≥ u , and on ∂B(x, r)\B(ξ, ρ) , w = uξ,ρ = u ,
w ≥ u on ∂B(x, r) , which by the maximum principle implies that w ≥ u in
B(x, r) . Apply the maximum principle once again to the domain B(x, r) ∩
B(ξ, ρ) and B(x, r)\B(ξ, ρ) , we conclude that w ≥ uξ,ρ in B(x, r) and hence
(4.23) holds.
Claim: Let u1 , · · · , uk ∈ σψ (Ω) , and v = max{u1 , · · · , uk } . Then v ∈

σψ (Ω) .
Proof. Given ξ ∈ Ω , for all sufficient small ρ such that B(ξ, ρ)⊂⊂Ω ,
v(ξ) = max{u1 (ξ), · · · , uk (ξ)} ≤ max{Mu1 (ξ, ρ), · · · , Muk (ξ, ρ)} ≤ Mv (ξ, ρ) .
Claim: For all given ψ ∈ C 0 (∂Ω) , the function wψ (x) ≡ sup w(x) is
w∈σψ (Ω)
well-defined and is harmonic in Ω .
Proof. wψ defined above is well-defined due to the maximum principle for

sub-harmonic functions, and is clearly in σψ (Ω) since
Z Z
sup − w(y)dSy ≤ − sup w(y)dSy .
w∈σψ (Ω) ∂B(ξ,ρ) ∂B(ξ,ρ) w∈σψ (Ω)
It suffices to show that wψ has the mean-value property. Suppose the con-
trary, then
Z
wψ (ξ) < − wψ (y)dSy
∂B(ξ,ρ)
for some ξ ∈ Ω and some ρ such that B(ξ, ρ)⊂⊂Ω . By the previous two
claims, the function (wψ )ξ,ρ belongs to σψ (Ω) , and wψ (ξ) < (wψ )ξ,ρ (ξ) .
Then wψ (ξ) 6= sup w(ξ) .
w∈σψ (Ω)
If (4.21) has a solution, it has to be equal to wψ defined above since the

solution itself belongs to σψ (Ω) . In other words, wψ is the only candidate
for the solution. In order to make sure that wψ solves (4.21), we need to
make sure that wψ satisfies the boundary condition.
Definition 4.34 (Barrier Property). A domain is said to have the barrier

property if for each η ∈ ∂Ω , there exists a function, called a barrier function,
Qη ∈ σ(Ω) ∩ C(Ω) for which
Qη (η) = 0 , Qη (x) < 0 for x ∈ ∂Ω , x 6= η .
The barrier property can be verified for a large class of domains Ω . For
example, if Ω is strictly convex in the sense that through each point η ∈ ∂Ω
there passes a hyperplane πη having only η in common with Ω , then Ω has
the barrier property.
As long as Ω has the barrier property, for each y ∈ ∂Ω , there exists a
sub-harmonic function w ∈ σψ (Ω) and w(y) = ψ(y) (for example, consider
w(x) = ψ(y) + Qy (x)). The only thing it remains to be proved is that
wψ ∈ C 0 (Ω) , or
lim wψ (x) = ψ(η) .

x∈Ω
x→η
Claim: If Ω has the barrier property, then for η ∈ ∂Ω ,
lim inf wψ (x) ≥ ψ(η) .

x∈Ω
x→η
Proof. For > 0 and K > 0 , the function v(x) = ψ(η) − + KQη (x) belongs
to σ(Ω) ∩ C 0 (Ω) , and satisfies
v(x) ≤ ψ(η) − ∀ x ∈ ∂Ω , v(η) = ψ(η) − .
Since ψ ∈ C 0 (∂Ω) , there exists δ > 0 such that |ψ(x) − ψ(η)| < whenever
|x − η| < δ , x ∈ ∂Ω ; thus v(x) ≤ ψ(x) if |x − η| < δ . If |x − η| ≥ δ , we
§4.5 Perron’s method and solutions to the Poisson problem 177
can choose K large enough so that v(x) ≤ ψ(x) since Qη has negative upper
bound on |x − η| ≥ δ . Therefore, v ∈ σψ (Ω) (if K is large enough). By the
definition of wψ , v(x) ≤ wψ (x) for all x ∈ Ω ; hence
ψ(η) − = lim inf v(x) ≤ lim inf wψ (η) . (4.24)

x∈Ω x∈Ω
x→η x→η
We then conclude the claim since (4.24) holds for all > 0 and all η ∈ ∂Ω .
Claim: If Ω has the barrier property, then for η ∈ ∂Ω ,
lim wψ (x) = ψ(η) .

x∈Ω
x→η
Proof. It suffices to show that
lim sup wψ (x) ≤ ψ(η) . (4.25)

x∈Ω
x→η
This is done by considering −w−ψ (x) which is defined in Ω by
−w−ψ (x) = − sup w(x) = inf v(x) .

w∈σ−ψ (Ω) −v∈σ−ψ (Ω)
For all w ∈ σψ (Ω) and −v ∈ σ−ψ (Ω) , w ≤ ψ ≤ v on ∂Ω , and w − v ∈

σ(Ω) ∩ C 0 (Ω) ; therefore by the maximum principle for the sub-harmonic
functions,
w≤v ∀x ∈ Ω ⇒ wψ (x) ≤ −w−ψ (x) ∀ x ∈ Ω .
By previous claim,
lim inf w−ψ (x) ≥ −ψ(η) ⇒ lim sup wψ (x) ≤ lim sup −w−ψ (x) ≤ ψ(η) .
x∈Ω x∈Ω x∈Ω
x→η x→η x→η
Theorem 4.35. If the domain Ω has the barrier property, and f ∈ C 0,α (Ω) ,
then there exists a unique solution u ∈ C 2 (Ω) ∩ C 0 (Ω) of the Dirichlet prob-
lem
−∆u = f in Ω
u=g on ∂Ω ,
for arbitrary continuous boundary value g .

4.6 Exercise
Problem 4.1. Let Ω = B(0, 12 ) ⊆ R2 denote the open ball of radius 1

2
centered at the origin. For x = (x1 , x2 ) ∈ Ω , let
" ! #
1
u(x1 , x2 ) = x1 x2 log log p 2 − log log 2 .
x1 + x22
(a) Show that u ∈ C 1 (Ω) ;
(b) Show that ∆u ∈ C(Ω) for j = 1, 2 , but that u 6∈ C 2 (Ω) .

2 (D) ∩ C(D) be a solution to the problem
Problem 4.2. Let u ∈ Cloc
−∆u = 1 in D ≡ (−1, 1) × (−1, 1) ⊆ R2 ,

u=0 on ∂D .
Show that u cannot belong to C 2 (D) .
Problem 4.3. Find a solution to the Dirichlet problem ∆u = 0 in the square
(x, y) ∈ R2 − 1 ≤ x ≤ 1 , −1 ≤ y ≤ 1

satisfying the boundary conditions

3π
u(x, y) = cos x , on y = ±1 , −1 ≤ x ≤ 1 ,
2
3π
u(x, y) = cos y , on x = ±1 , −1 ≤ y ≤ 1 .
2
Problem 4.4. Let u , v be smooth harmonic functions, such that
u(tx) ≡ ta u(x) , v(tx) = tb v(x)
for all x ∈ Rn , t > 0 , with constants a 6= b . Use Green’s identity to show

that
Z
uvdS = 0 .
∂B(0,1)
Problem 4.5. Let u be a harmonic function in the unit ball B1 ≡ B(x0 , 1) ⊂

Rn . Prove the following gradient estimate:
h i
|Du(x0 )| ≤ n sup u(x) − u(x0 )
x∈B1
§4.6 Exercise 179
Hint: Note that all the derivatives ∂xi u are harmonic in B1 , so that by the
mean value and divergence theorems,
∂u 1 ∂u 1
Z Z
(x0 ) = (x)dx = uNi dS ,
∂xi |B1 | B1 ∂xi |B1 | ∂B1
where Ni is the i-th component of the unit normal N to ∂B1 . Obviously, for
x ∈ ∂B1 , we have N(x) = x − x0 .
2 (R2 ) ∩ C (R2 )
Problem 4.6. Show that there are no functions u ∈ Cloc + loc +
satisfying the properties
u ≥ 0, ∆u = 0 in R2+ ≡ {x = (x1 , x2 ) ∈ R2 | x2 > 0} , u(x1 , 0) = x21 .
Hint: Suppose there exists such a function u . For arbitrary R > 0 , compare
2 (R2 ) ∩ C(R2 ) to the problem
u with the solution v ∈ Cloc + +
∆v = 0 in R2+ , v(x1 , 0) = ζ(x1 )x21 ,
where
ζ ∈ Cc∞ (−2R, 2R) , 0 ≤ ζ ≤ 1, and ζ ≡ 1 on [−R, R] .
Problem 4.7. Let u(x1 , x2 ) ∈ C 2 (Ω) , where Ω ≡ {x = (x1 , x2 ) ∈ R2 | x1 >

0 , x2 > 0} , such that
∆u = 0 in Ω, u=0 on ∂Ω .
In addition, let |u(x)| ≤ c1 +c2 |x| in Ω with some constants c1 and c2 . Show
that u ≡ 0 in Ω .
Hint: Extend the domain where u is defined to R2 and use the gradient
estimates to show that u is in fact a linear function.
Problem 4.8. Let u(x) = u(x1 , x2 ) be a bounded solution of the Laplace

equation
∆u = 0 in R2+
with boundary condition

|x1 |
u(x1 , 0) = g(x1 ) = .
1 + x21
Show that the gradient Du is unbounded on R2+ .

Problem 4.9. Find the Green function for the domain Ω = x ∈ Rn |x|2 <

1 , xn > 0 and the corresponding Green’s representation formula for the
solution of the Dirichlet problem
∆u = 0 in Ω,
u=f on ∂Ω .
Chapter 5
Second Order Linear Elliptic

Equations
5.1 The strong form of 2nd-order elliptic bound-

ary value problems
Let Ω ⊆ Rn denote an open and smooth domain, and suppose that u ∈

C 2 (Ω). We consider the following second-order differential operators defined
in Ω:
∂ ij ∂u ∂u
(Lu)(x) = − a (x) (x) + bi (x) (x) + c(x)u(x)
∂xi ∂xj ∂xi

= − div a(x) ·Du(x) + b(x) ·Du(x) + c(x)u(x) (divergence form)
or
∂2u ∂u
(Lu)(x) = −aij (x) (x) + bi (x) (x) + c(x)u(x)
∂xi ∂xj ∂xi
= − a(x) :D2 u(x) + b(x) ·Du(x) + c(x)u(x) (non-divergence form).
We remind the reader, that we are employing the Einstein summation con-
vention, in which repeated Latin indices are summed from 1 to n, and re-
peated Greek indices are summed from 1 to n − 1. For example, fi gi =
Pn n−1
P
fi gi and fα gα = fα gα . Also, our intrinsic notation, is to use dots to
i=1 i=1
indicate contraction between indices, so that the scalar inner-product of two
vectors X and Y is denoted by X · Y = Xi Yi , and the trace of the product
of two matrices A and B is written as A : B = Aij Bji .
181
182 CHAPTER 5. Second Order Linear Elliptic Equations
Definition 5.1. The operator L is said to be elliptic at a point x ∈ Ω if

there exist positive functions λ(x) and Λ(x) such that
0 < λ(x)|ξ|2 ≤ aij (x)ξj ξi ≤ Λ(x)|ξ|2 (5.1)
for all ξ = (ξ1 , · · · , ξn ) ∈ Rn − {0}.

If there exists positive constants λ 0 and Λ0 such that λ(x) ≥ λ 0 > 0 and
Λ(x) ≤ Λ0 < ∞, then L is uniformly elliptic in Ω.
Definition 5.2 (Elliptic Boundary Value Problem). Given an (uniformly)

elliptic operator L, we say that u : Ω → R is a strong solution of an elliptic
boundary value problem, or BVP, if
(a) Lu(x) = f (x) for almost all x ∈ Ω;
(b) B∂Ω u(x) = g(x) for almost all x ∈ ∂Ω,
where B∂Ω denotes a “boundary” operator, f is the interior forcing function,

and g is the prescribed boundary condition.
While it is possible to impose a wide variety of boundary conditions,

perhaps the most fundamental are known as the Dirichlet, Neumann, and
periodic boundary conditions. Here is a partial list of boundary conditions
that we shall examine:
1. B∂Ω u = u on ∂Ω (Dirichlet boundary condition);

∂u
2. B∂Ω u = aij Ni on ∂Ω (Neumann boundary condition);
∂xj
3. with ∂Ω = Γ1 ∪ Γ2 so that Γ1 ∩ Γ2 = ∅, we set
∂u
B∂Ω u = 1Γ1 u + 1Γ2 aij Ni (Mixed boundary conditions);
∂xj
∂u
4. B∂Ω u = aij Ni + ku on ∂Ω (Robin boundary condition);
∂xj
5. with g(x) = 0 and Ω = Tn , and with ei denoting the n orthonormal
unit vectors,
B∂Ω u(x) = u(x + ei ) − u(x) (Periodic boundary conditions);
We will additionally describe certain non standard boundary conditions,

involving elliptic differential operators on the boundary ∂Ω.
5.1. STRONG SOLUTIONS 183
5.1.1 The Dirichlet problem
Given an interior forcing function f : Ω → R and a function g : ∂Ω → R

on the boundary, the solution to the Dirichlet problem on Ω is a function
u : Ω → R which satisfies
Lu = f in Ω, (5.2a)
u=g on ∂Ω . (5.2b)
The Dirichlet boundary condition is also known as an essential boundary

condition. This is due to the inclusion of the boundary condition into the
functional framework used for existence theorems (as will be made precise
below).
5.1.2 The Neumann problem
Given an interior forcing function f : Ω → R and a function g : ∂Ω → R

on the boundary, the solution to the Neumann problem on Ω is a function
Lu = f in Ω, (5.3a)
∂u
aij Ni = g on ∂Ω . (5.3b)
∂xj
The Neumann boundary condition is also known as a natural boundary con-

dition. Unlike the Dirichlet boundary condition, this boundary condition
appears naturally in the variational form of the boundary value problem.
5.1.3 The Robin problem
It is also to possible to impose certain linear combinations of the Dirichlet

and Neumann boundary conditions. In this case, given an interior forcing
function f : Ω → R and a function g : ∂Ω → R on the boundary, the solution
to the Robin problem on Ω is a function u : Ω → R which satisfies
Lu = f in Ω, (5.4a)
∂u
aij Ni + ku = g on ∂Ω . (5.4b)
∂xj
The sign of k will play a crucial role in existence theorems.

5.1.4 A non-standard elliptic problem
Motivated by problems involving surface tension (and more generally, surfac-

tant theory), we shall also analyze a non standard boundary value problem,
in which a Neumann-like boundary operator is supplemented by a “surface”
operator (to be read as boundary operator) which is itself elliptic.
A prototype BVP is the following: Given an interior forcing function
f : Ω → R and a function g : ∂Ω → R on the boundary, find a function
Lu = f in Ω, (5.5a)
∂u
aij Ni = ∆g u + g on ∂Ω . (5.5b)
∂xj
In this setting, the smooth boundary ∂Ω is to be thought of as (n − 1)-

dimensional Riemannian manifold1 with Riemannian metric g: (∂Ω, g). The
Riemannian metric is a positive definite (n − 1) × (n − 1) matrix which in a
local coordinate basis is denoted by gαβ , where the indices α, β = 1, ..., n−1.
The inverse of the metric g −1 is denoted with raised superscripts g αβ .
The operator ∆g is called the Laplace-Beltrami operator on ∂Ω is defined
as
1 ∂ p ∂u
∆g u = p det(g)g αβ . (5.6)
det(g) ∂xα ∂xβ
The Laplace-Beltrami operator (5.6) is the natural generalization of the

Laplacian on Rn−1 .
5.1.5 A vector-valued Dirichlet problem: Linear Elasticity
Our examples, thus far, have all been for scalar-valued solutions, but the
majority of problems coming from physics and engineering are indeed vector-
valued. Linear elasticity affords a nice example of a second-order boundary
value problem for a vector valued unknown. The so-called displacement
1
When Ω is a connected and simply connected bounded subset of Rn , then for the case
that n = 2, the boundary ∂Ω is a closed curve and in the case that n = 3, the boundary ∂Ω
is a closed surface. For the former, the reader may suppose that ∂Ω is diffeomorphic to the
circle S1 , while for the latter, the boundary ∂Ω can be thought of as diffeomorphic to the
unit sphere S2 . While both S1 and S2 are particular examples of Riemannian manifolds,
standard polar coordinates can be employed to parameterize such boundaries.
5.1. STRONG SOLUTIONS 185
vector u : Ω ⊆ R3 → R3 , measures the displacement of an elastic material

in the three physical space dimensions. In particular,
u(x) = (u1 (x), u2 (x), u3 (x)) ,
where u1 and u2 capture the amount of in-plane (horizontal) deformation,

and the u3 represents the vertical (or out of plane) deformation of the ma-
terial.
In order to describe the associated BVP problem, we must generalize
the setting of the operator L from possessing a coefficient matrix Aij (which
depends on two indices) to a coefficient tensor C ijk` (which depends on four
indices). In the setting of elasticity, such a tensor encodes the material
response to tension, compression, shear forcing, moments, etc., in all of the
three space directions. In the parlance of continuum mechanics, C encodes
Hooke’s Law relating the strain of an elastic material to the resulting stress.
Definition 5.3 (The fourth-rank tensor C). The fourth-rank tensor C, writ-
ten as C ijk` with respect to the standard coordinate basis in R3 , satisfies the
following properties:
1. Positivity: ∃λ > 0, C ijk` ξi ξj ηk η` ≥ λ|ξ|2 |η|2 for all ξ, η ∈ R3 .
2. Symmetry: C ijk` = C jik` = C ij`k = C kìj for all i, j, k, ` = 1, 2, 3.
Definition 5.4 (Strain and stress). The strain tensor is defined to by the
symmetric part of the 3 × 3-matrix Du. In components,
∂uk ∂u`
kl = + .
∂x` ∂xk
The stress tensor σ is then defined as
σ ij = C ijk` k` .
∂uk
Due to the symmetry assumption on C ijk` , we see that σ ij = C ijk` .
∂x`
Definition 5.5 (Divergence of a matrix). Suppose that M ij (x) denotes a
differentiable matrix function on Ω. The divergence of M , is the n-vector
given by
∂M ij
[Div M ]i = .
∂xj
The equations of linear elasticity are the static conservation law
Div σ = f in Ω
for a vector forcing function f : Ω → R3 . More explicitly, we have that
∂ ijk` ∂u`
− C = fi in Ω, (5.7a)
∂xj ∂xk
(B∂Ω u)i = g i on ∂Ω , (5.7b)
where f i is the interior forcing vector and g i denotes the vector of boundary
data. The operator B can be chosen to represent either the vector Dirichlet,
Neumann, mixed-type boundary conditions, or periodic boundary conditions
can be imposed if the domain Ω = T3 .
Of particular interest to us will be the specification of Dirichlet and
Neumann boundary conditions:
1. [B∂Ω u]i = ui on ∂Ω (Dirichlet boundary condition);
∂uk
2. [B∂Ω u]i = C ijk` Nj on ∂Ω (Neumann boundary condition).
∂x`
Note that prescribing the Dirichlet conditions on ∂Ω requires us to prescribe
the displacement field ui of the elastic material on the boundary ∂Ω. On the
other hand, the Neumann boundary condition, also known as the traction
boundary condition, requires us to instead specify, the so-called traction
vector, defined to be the normal component of the stress tensor, σ ij Nj , on
the boundary ∂Ω.
5.2 Variational Formulations of Elliptic BVPs
The strong formulation of second-order elliptic BVPs as given by Definition

5.2 requires that linear combinations of partial derivatives of the solution
should equal a given forcing function at almost every point. Suppose that
the solution u(x) represents the equilibrium state of the temperature dis-
tribution. Is it possible to measure the temperature at every point of the
continuum, and moreover, is it possible to produce a change in the tempera-
ture (via forced heating) at every point? If measurements of the temperature
are made with a sensor of some sort, then there exists a smallest physical
§5.2 Variational formulations of elliptic BVPs 187
scale, below which it is impossible to discern temperature differences. In par-

ticular, the sensor “averages” about a small region A (with Zdiameter below
the smallest physical scale), and in fact produces the result u(x)φ(x) dx,
A
where φ(x) depends on the type of sensor that is employed.
This motivates to consider weighted averages of elliptic BVPs, and per-
haps the simplest case to consider is the Dirichlet BVP with homogeneous
boundary condition u = 0. To this end, suppose that φ ∈ C0∞ (Ω), and
consider instead of Lu = f in Ω and u = 0 on ∂Ω, the following integrated
equality:
Z Z
Lu(x)φ(x) dx = f (x)φ(x) dx ∀ φ ∈ C0∞ (Ω) . (5.8)
Ω Ω
Clearly, if (5.8) holds, then Lu(x) = f (x) for a.e. x ∈ Ω. On the other hand,
the introduction of the integral over the domain Ω allows us to employ one
of the most powerful tools in analysis: integration by parts. We can rewrite
(5.8) as
Z Z Z
ij i
a (x)u,j (x)φ,i(x) dx + b (x)u,i(x)φ(x) dx + c(x)u(x)φ(x) dx
Ω Z Ω Z Ω
− aij (x)u,j (x)Ni (x)φ(x)dSx = f (x)φ(x) dx ∀ φ ∈ C0∞ (Ω) ,

∂Ω Ω
(5.9)
and as the test function φ vanishes on ∂Ω, the boundary integral in (5.9) is
equal to zero, so that
Z Z Z
ij i
a (x)u,j (x)φ,i(x) dx + b (x)u,i(x)φ(x) dx + c(x)u(x)φ(x) dx
Ω Z Ω Ω
= f (x)φ(x) dx ∀ φ ∈ C0∞ (Ω) . (5.10)

Ω
As the coefficient matrix aij ∈ L∞ (Ω), and as our test function φ has square-
integrable partial derivatives, we see that the integral equality (5.10) is well
defined whenever both u and u,j are square-integrable, i.e., u ∈ H 1 (Ω).
Thus, we have reduced the regularity requirements of the solution u,
from C 2 (Ω) to H 1 (Ω). As we will explain below, the integral equality given
in (5.10) (with certain functional constraints) is known as the variational
formulation of the Dirichlet problem. In order for solutions of such a vari-
ational problem to satisfy the homogeneous Dirichlet boundary condition
u = 0, we insist that u be element of a subspace of H 1 (Ω); namely, we

require solutions u of (5.10) to be in the subspace H01 (Ω). Furthermore, we
as C0∞ (Ω) is dense in H01 (Ω), we may consider all test functions to be in
H01 (Ω) as well.
With H01 (Ω) acting as both the space of solutions and the space of test
functions, we have converted a second-order elliptic BVP to a variational
problem in a Hilbert space, a setting very much amenable to the Lax-
Milgram theorem for the existence and uniqueness of solutions. In turn,
the Lax-Milgram theorem is a generalization of the Riesz Representation
theorem.
To explain the choice of the spaces V used for the various boundary
conditions discussed above, it will be convenient to first examine a simpli-
fied version of the operator L in which the lower-order terms vanish. In
particular, we shall assume2 that

∂ ij ∂
L=− a (x) .
∂xi ∂xj
Definition 5.6 (Bilinear form associated to L). We let B : V × V → R be
given by
∂u ∂v
Z
B(u, v) = aij (x) dx u, v ∈ V ,
Ω ∂xj ∂xi
where V denotes a subspace of H 1 (Ω).
Definition 5.7 (Weak Solutions to elliptic BVP). Given f ∈ L2 (Ω), say

that u ∈ V is a weak solution of Lu = f in Ω with B∂Ω u = 0 on ∂Ω if
Z
B(u, v) = f (x)v(x) dx ∀ v ∈ V . (5.11)
Ω
The choice of subspace V will depend on the type of homogeneous boundary
condition that is imposed. For certain boundary conditions, (5.11) will be
supplemented by an additional bilinear form b : v × v → R, where v denotes
a subspace of H 1 (∂Ω) and will be specified below. If we allow the possibility
that b = 0, then a more general form of (5.11) can be written as follows:
Find u ∈ V ∩ v such that
Z
B(u, v) + b(u, v) = f (x)v(x) dx ∀ v ∈ V ∩ v . (5.12)
Ω
2
We will return to the general operator L below.
§5.2 Variational formulations of elliptic BVPs 189
We begin the discussion of the Dirichlet problem (5.2) with homogeneous

boundary conditions u = 0. Later, we will explain how to convert the general
problem to the homogeneous case.
As we described above, in this case, we set V = H01 (Ω).
Definition 5.8. We say that u ∈ H01 (Ω) is a weak solution to the homoge-
neous Dirichlet problem if
Z
B(u, v) = f (x)v(x) dx ∀ v ∈ H01 (Ω) .
Ω
Notice that the null space of L with u = 0 boundary conditions is trivial;

on the contrary, the null space of L with homogeneous Neumann boundary
conditions is one-dimensional and is spanned by the constant function u = 1.
One of the fundamental ideas in writing the variational form of the BVP
is to choose the functional framework in such a way as to mod-out any
nontrivial kernel. This “modding-out” coincides with the solvability criteria
of the Fredholm Alternative Theorem that we discuss in Section 5.7.
As such, for the homogeneous Neumann problem, we set V = H 1 (Ω)/R.
Definition 5.9. We say that u ∈ H 1 (Ω)/R is a weak solution to the Neu-

mann problem if
Z Z
B(u, v) = f (x)v(x) dx + g(x)v(x) dSx ∀ v ∈ H 1 (Ω)/R .
Ω ∂Ω
For the homogeneous Robin problem, we observe that the kernel of L is once
again trivial. In this case we set V = H 1 (Ω).
Definition 5.10. We say that u ∈ H 1 (Ω) is a weak solution to the homo-

geneous Robin problem if
Z Z Z
B(u, v) + k(x)u(x)v(x) dSx = f (x)v(x) dx + g(x)v(x) dx
∂Ω Ω ∂Ω
∀ v ∈ H 1 (Ω) .
5.2.4 A non-standard elliptic problem
We next consider the variational form of the elliptic BVP Lu = f in Ω with

the non-standard boundary condition aij u,j Ni − ∆g u = 0 on ∂Ω.
Following Definition 5.10, we see that integration by parts leads to the
following variational form
Z
B(u, v) − (∆g u)v dSx = (f, v)L2 (Ω) (5.13)
∂Ω
for all v in some space of test functions. Just as we used integration by parts
on the interior
Z elliptic operator, we can now do the same on the boundary
integral (∆g u)v dSx . In order to do so, we make use of a local coordinate
∂Ω
system (x1 , ..., xn−1 ) for ∂Ω, and note the surface measure is given by
p
dS = det g dx1 · · · dxn−1 .
It follows that
Z p Z
− (∆g u)v dSx = − [ det gg αβ u,β ],α v dx1 · · · dxn−1
∂Ω Z ∂Ω p
= g αβ u,β v,α det g dx1 · · · dxn−1
Z∂Ω

= g gradu(x), gradv(x) dSx ,
∂Ω
where gradu denotes the surface gradient of the function u. It follows that
we require both H 1 (Ω) regularity on the interior as well as H 1 (∂Ω) on the
boundary. Note that the kernel of L is once again spanned by the constant
functions.

This leads us to set V = H 1 (Ω)/R ∩ H 1 (∂Ω) = w ∈ H 1 (Ω)/R w ∈

H 1 (∂Ω) . We make the following

Definition 5.11. We say that u ∈ H 1 (Ω)/R ∩ H 1 (∂Ω) is a weak solution

to the non standard elliptic BVP if
Z Z

B(u, v) + g gradu(x), gradv(x) dSx = f (x)v(x) dx
∂Ω Ω
∀ v ∈ H 1 (Ω)/R ∩ H 1 (∂Ω) .
5.3. THE LAX-MILGRAM THEOREM 191
5.2.5 Linear elasticity

Dirichlet boundary conditions
The case of homogeneous Dirichlet boundary conditions for the vector prob-
lem is the same as for the scalar problem. In particular, we set V =
H01 (Ω; R3 ).
Definition 5.12. We say that u ∈ H01 (Ω; R3 ) is a weak solution to the

system of equations of linear elasticity with homogeneous Dirichlet boundary
conditions if
∂u` ∂v i
Z Z
C ijk` dx = f (x) · v(x) dx ∀ v ∈ H01 (Ω; R3 ) .
Ω ∂xk ∂xj Ω
Neumann boundary conditions
When we prescribe the traction vector on the boundary rather than the
displacement vector, the equation Div σ = 0 has a six-dimensional null space
spanned by the translations and rotations of R3 . In particular, the kernel is
spanned by α1 + α2 x where α1 is any constant vector and α2 is any skew-
symmetric matrix. Let S denote this span. Then we set V = H 1 (Ω; R3 )/S.
Definition 5.13. We say that u ∈ H 1 (Ω; R3 )/S is a weak solution to the

system of equations of linear elasticity with homogeneous Neumann boundary
conditions if
` i
ijk` ∂u ∂v
Z Z
C dx = f (x) · v(x) dx ∀ v ∈ H 1 (Ω; R3 )/S .
Ω ∂xk ∂xj Ω
5.3 The Lax-Milgram theorem
A general framework for proving the existence of a unique weak solution is

based on a generalization of the Riesz Representation theorem.
The objective is to find a unique weak solution given by Definition 5.7.
We restate our definition as follows: we suppose that V is a reflexive Banach
space, and that we have a continuous bilinear map B : V × V → R for which
we search for a unique element u ∈ V such that B(u, v) = hf, vi for all v ∈ V.
We are using the notation h·, ·i to denote the duality pairing between V and
its dual space V 0 , and we assume that the interior forcing function induces
a continuous linear functional on V. For second-order elliptic BVPs, the

Banach space V will be a subspace of the Hilbert space H 1 (Ω) (or H 1 (Ω; Rn )
in the case of vector-valued solutions), and forcing function function will
typically have L2 (Ω) regularity.
The bilinear form B induces a linear operator L : V → V 0 defined by
Lu = B(u, ·) ∀u ∈ V or hLu, vi = B(u, v) ∀ u, v ∈ V .
Therefore, the variational formulation can be stated as
Lu = F ∈ V 0 ,
which means that

hLu, viV = hF, viV ∀v ∈ V.
Let B(V, V 0 ) denote the space of bounded linear operators from V to V0 .
Definition 5.14. L ∈ B(V, V 0 ) is coercive if there exists λ 0 > 0 such that
λ 0 kuk2V ≤ hLu, uiV ∀u ∈ V,
Theorem 5.15 (Lax-Milgram). If L ∈ B(V, V 0 ) and is coercive, then there

exists a unique solution to Lu = f ∈ V 0 and u ∈ V satisfies
1
kukV ≤ kf kV 0 . (5.14)
λ0
Proof. Step 1: (Uniqueness) Assume that v is also a solution to Lv = f .

Then
ku − vk2V ≤ λ−1
0 hL(u − v), u − viV = 0 .
Hence u = v in V.
Step 2: (Estimate) Assume that u ∈ V satisfies Lu = f ∈ V 0 . Then
kuk2V ≤ λ−1 −1 −1
0 hLu, uiV = λ 0 hf, uiV ≤ λ 0 kf kV 0 kukV ;
thus kukV ≤ λ−1

0 kf kV 0 .
Step 3: (Existence)
5.4. THE BBL INF-SUP CONDITION 193
1. We first prove that the range of L, R(L), is closed. Since R(L) is a

metric space, it suffices to show R(L) is complete. Let yn ∈ R(L) be
a Cauchy sequence; that is, given > 0, there exists N > 0 so that
kyn − ym kV 0 ≤ whenever n, m ≥ N . Since each yn ∈ R(L), there
exists xn ∈ V such that Lxn = yn . Since L(xn − xm ) = yn − ym , by
(5.14),
kxn − xm kV ≤ λ−1
0 ∀ n, m ≥ N .
This implies xn is a Cauchy sequence in V which is a Banach space so

∃ x ∈ V such that xn → x ∈ V. Now
lim yn = lim Lxn = L lim xn = Lx ≡ y ∈ R(L) .

n→∞ n→∞ n→∞
So R(L) is closed.
2. Next, we show that R(L) = V 0 . Assume to the contrary that R(L) ⊆

V 0 . By the Hahn-Banach theorem, ∃ u ∈ (V 0 )0 = V (by reflexivity)
such that u|R(L) = 0 but hu, f i = 1. Then
kuk2V ≤ λ−1
0 hLu, uiV = 0
which contradicts the conclusion that u = 0.
5.4 The BBL inf-sup condition3
The Lax-Milgram theorem is not sufficient to cover certain elliptic systems

of mixed type, such as a formulation of the Stokes equations with explicit
pressure and velocity. We shall describe a theory which generalizes the
Lax-Milgram theorem without assuming that L maps V into its dual space
V 0 . The replacement for the coercivity is the so-called inf-sup condition
introduced by Babuška, Brezzi, and Ladyzhenskaya.
Let V and W be Hilbert spaces with inner products (·, ·)V and (·, ·)W .
Let B : V × W → R be bilinear and continuous:
B(v, w) ≤ kBkkvkV kwkW ∀v ∈ V,w ∈ W.

3
Readers should skip this section on first reading.
In particular, we could have V = W. Consider L : V → W 0 , L∗ : W → V 0

defined by
hLv, wiW = hL∗ w, viV = B(v, w) ∀v ∈ V,w ∈ W.
So L∗ is the adjoint operator associated with L. Note that by the continuity

of B, L and L∗ are both continuous.
Theorem 5.16 (Inf-Sup). The linear map L : V → W 0 is an isomorphism

if and only if the following conditions hold:
B(v, w)
(1) there exists α > 0 so that sup ≥ αkvkV ,
w∈W kwkW
(2) for every non-zero w ∈ W, there exists v ∈ V so that B(v, w) 6= 0.
Remark 5.17. The condition (1) is equivalent to
B(v, w)
inf sup ≥ α > 0. (5.15)
v∈V w∈W kvkV kwkW
and the condition (2) is equivalent to
sup B(v, w) > 0 ∀ w 6= 0 . (5.16)

v∈V
Note that (5.16) is also equivalent to ker L∗ = {0}.
Lemma 5.18 (The closed range theorem). Let X and Y be Banach spaces,
and L : X → Y be a continuous linear map. Then the following statements
are equivalent:
1. The image L(X) of X is closed in Y,
2. L(X) = (ker L∗ )0 , where for a closed subspace A ⊆ V∗ , A0 is the polar

set of A defined by
A0 ≡ w∗ ∈ V 0 hw∗ , ai = 0 ∀ a ∈ A .

Lemma 5.19 (Brezzi). The following statements are equivalent.

B(v, w)
(a) there exists α > 0 so that sup ≥ αkvkV ,
w∈W kwkW
§5.4 The BBL inf-sup condition 195
(b) L : V → W00 is an isomorphism and
kLvkW 0 ≥ αkvkV ∀v ∈ V,
(c) L∗ : W0⊥ → V 0 is an isomorphism and
kL∗ wkV 0 ≥ αkwkW ∀ w ∈ W⊥

0 ,
where W0 ≡ w ∈ W B(v, w) = 0 ∀ v ∈ V = ker L∗ .

Proof. Step 1: (a) ⇒ (b)

hLv, wi B(v, w)
kLvkW 0 = sup = sup ≥ αkvkV .
w∈W kwk W w∈W kwkW
This implies that L is one-to-one and so invertible in R(L) = L(V) which is

closed (same argument as the proof of the Lax-Milgram Theorem). More-
over, L−1 : R(L) → V is continuous. By the closed range theorem, L : V →
W00 is an isomorphism.
Step 2: (b) ⇒ (c)
Let w ∈ W⊥ ∗ 0
0 be arbitrary and consider map w ∈ W , z 7→ (z, w)W =
hw∗ , zi = w∗ (z). Since (z, w)W = 0 for all z ∈ W0 , we see that w∗ ∈ W00 .
By (i), there is a unique v ∈ W so that Lv = w∗ . Then
αkvkV ≤ kLvkW 0 = kw∗ kW 0 = kwkW .
Therefore,
hw∗ , wi B(v, w) hL∗ w, vi hL∗ w, vi
αkwkW = α =α =α ≤ .
kwkW kwkW kwkW kvkV
This implies L∗ is one-to-one, (L∗ )−1 : L∗ (W⊥ ⊥
0 ) → W0 is continuous,
and R(L∗ ) is closed. We then conclude by the closed graph theorem that
L∗ (W⊥ 0 0 0
0 ) = (ker L) = {0} = V .
Step 3: (c) ⇒ (a)
Take an arbitrary v ∈ V, and let w ∈ W⊥ ∗ ∗ 0
0 satisfy L w = v ∈ V with
kL∗ wkV 0 ≥ αkwkW . This implies
hv ∗ , vi hL∗ w, vi B(v, w)
kvkV = ∗
= ∗
≤ .
kv kV 0 kL wkV 0 αkwkW
Therefore,
B(v, w) B(v, w) B(v, w)
sup ≥ sup = sup ≥ αkvkV .
w∈W kwk W w∈W⊥ kwk W v ∈V kwkW
∗ 0
0
Proof of Theorem 5.16.
(⇐) Assume (1) and (2). Since (1) is the same as (a) of Lemma 5.19, (1)
⇔ (b) so that L : V → W00 is an isomorphism. It remains to prove
that W00 = W 0 . By (2), W0 = ker L∗ = {0} and hence W00 = W 0 .
(⇒) Suppose L : V → W 0 is an isomorphism. Given v ∈ V, let w∗ = Lv.

Since L−1 : W 0 → V is bounded, we find that
hLv, wi 1 1
sup = kLvkW 0 = kw∗ kW 0 ≥ kL−1 w∗ kV = kvkV .
w∈W kwkW C C
Therefore, α = 1/C will do the job for (1).
Moreover, L(V) = W 0 = (ker L∗ )0 , whence ker L∗ = {0} which implies

(2).
Theorem 5.20 (Inf-Sup, another form). The following inf-sup condition is

equivalent to (1) and (2) in Theorem 5.16:
B(v, w) B(v, w)
inf sup = inf sup > 0. (5.17)
v∈V w∈W kvkV kwkW w∈V v∈W kvkV kwkW
B(v, w) B(v, w)
Proof. Let α = inf sup and β = inf sup . We first
kvkV kwkW
v∈V w∈W w∈W v∈V kvkV kwkW
show that (1) and (2) are equivalent to α > 0 and β > 0, and then prove
α = β.
(⇐) Note that (1) is the same as α > 0. It suffices to prove (1) and
(2) implies β > 0. Since (2) implies W0 = ker L∗ = {0}, W⊥
0 =
W. (5) (which is equivalent to (1)) then implies L∗ : W → V 0 is an
isomorphism with kL∗ wkV 0 ≥ αkwkW . Therefore,
B(v, w) 1 hL∗ w, vi 1
sup = sup = kL∗ wkV 0 ≥ α .
v∈V kvkV kwkW kwkW v∈V kvkV kwkW
(⇒) Suppose that α and β are both positive. β > 0 implies that ker L∗ =
{0} which is equivalent to (2). (1) is again the same as α > 0.
By Brezzi’s lemma, kL∗ wkV 0 ≥ αkwkW . Therefore,

B(v, w)
sup ≥α
v∈V kvkV kwkW
which implies β ≥ α. Since the argument is symmetric, α ≥ β as well and
hence α = β.
5.5. EXISTENCE AND UNIQUENESS OF WEAK SOLUTIONS 197
Remark 5.21. We can write the condition α = β in a different form

kL−1 k = k(L∗ )−1 k for
1 1 kLvkW 0 B(v, w)
= = inf = inf sup
kL−1 k kvkV v∈V kvkV v∈V w∈W kvkV kwkW
sup
v∈V kLvk W 0
B(v, w) 1
= inf sup = .
w∈W v∈V kvkV kwkW k(L )−1 k
∗
5.5 Existence and uniqueness of weak solutions
Basic Assumptions. In order to highlight the fundamental concepts, we

shall restrict our attention in this section to the case in which all of the lower-
order terms appearing in the operator L vanish: bi (x) = 0 for i = 1, ..., n
and c(x) = 0. In particular, we assume the following:

∂ ij ∂
1. L = − a (x) ;
∂xi ∂xj
2. L is uniformly elliptic in Ω;
3. the interior forcing function f is in L2 (Ω);
4. the boundary forcing g = 0 on ∂Ω.
The bilinear form B : V × V → R reduces to

∂u ∂v
Z
B(u, v) = aij (x) dx ,
Ω ∂xj ∂xi
and L : V ∩ v → (V ∩ v)0 is given by
hLu, vi = B(u, v) + b(u, v) ∀ u, v ∈ V ∩ v . (5.18)
5.5.1 A few Preliminary Lemmas
Lemma 5.22 (Poincaré inequality). Let Ω ⊆ Rn denote an open, bounded,

connected, and smooth domain. Then
ku − ūkL2 (Ω) ≤ CkDukL2 (Ω) ∀ u ∈ H 1 (Ω) , (5.19)

Z
where ū := − u(y)dy denotes the average value of u over Ω and the constant
Ω
C depends on Ω.
Proof. Suppose for the sake of contradiction that (5.22) does not hold. Then
there is a sequence {uj }∞ 1
j=1 ⊆ H (Ω) satisfying
kuj − ūj kL2 (Ω) > jkDuj kL2 (Ω) , (5.20)
with an associated sequence on the unit ball of H 1 (Ω) given by

uj − ūj
wj = with kwj kL2 (Ω) = 1 and w̄j = 0 .
kuj − ūj kL2 (Ω)
According to (5.20), kDwj kL2 (Ω) < j −1 , so that kwj k2H 1 (Ω) < 1 + j −2 <
∞. Strong compactness, given by Theorem 2.109 (see also Theorem 2.176)
provides a subsequence {wj k } and a limit w ∈ L2 (Ω) such that wj k → w in
L2 (Ω) as k → ∞. The limit w satisfies w̄ = 0 and kwkL2 (Ω) = 1.
Letting φ ∈ C0∞ (Ω). We see that
Z Z
w(x)Dφ(x) dx = lim wj k (x)Dφ(x) dx
Ω k→∞ Ω
Z
= − lim Dwj k (x) φ(x) dx ≤ lim jk−1 kφkL2 (Ω) = 0 .
k→∞ Ω k→∞
This shows that the weak derivative of w exists and is equal to zero almost
everywhere, i.e. w ∈ H 1 (Ω) and Dw = 0 a.e. As Ω is connected, we see
that w is a constant, and since w̄ = 0, we see that w = 0, contradicting the
fact that kwkL2 (Ω) = 1.
The identical proof also shows that the following two lemmas hold:
Lemma 5.23 (Poincaré inequality for H01 (Ω)). Let Ω ⊆ Rn denote an open,
bounded, connected, and smooth domain. Then
kukL2 (Ω) ≤ CkDukL2 (Ω) ∀ u ∈ H01 (Ω) , (5.21)
where the constant C depends on Ω.
Lemma 5.24 (Poincaré inequality for the Robin problem). Let Ω ⊆ Rn

denote an open, bounded, connected, and smooth domain. Then for k ∈
L∞ (∂Ω) and k ≥ 0 on ∂Ω and k > 0 on a set of surface measure greater
than zero. Then
√
kukL2 (Ω) ≤ C k k uk2L2 (∂Ω) + kDukL2 (Ω) ∀ u ∈ H 1 (Ω) , (5.22)
where the constant C depends on Ω.

§5.5 Existence and uniqueness of weak solutions 199
Corollary 5.25. Whenever ū = 0, kDukL2 (Ω) is an equivalent norm on

H 1 (Ω). In particular, there exists constants C1 , C2 such that
C1 kDukL2 (Ω) ≤ kukH 1 (Ω) ≤ C2 kDukL2 (Ω) .
According to Corollory 5.25, kDukL2 (Ω) is an H 1 (Ω)-equivalent norm

on the spaces H01 (Ω) and H 1 (Ω)/R. Of course, the Poincaré inequality
also holds for vector-valued functions as well. Another important inequality
follows:
Lemma 5.26 (Korn’s first inequality). Suppose that Ω satisfies the assump-
tions of the Poincaré inequality. Then
kukH 1 (Ω) ≤ CkkL2 (Ω) ∀ u ∈ H01 (Ω; R3 ) ,
1 ∂ui ∂uj
where ij = ( + ) denotes the symmetric part of Du.
2 ∂xj ∂xi
Lemma 5.27 (Korn’s second inequality). Suppose that Ω satisfies the as-
sumptions of the Poincaré inequality. Then
kukH 1 (Ω) ≤ CkkL2 (Ω) ∀ u ∈ H 1 (Ω; Rn )/S .
The proofs of these two lemmas are similar to the proof the Poincaré
inequality, and are left to the reader.
Remark 5.28. The Korn inequalities show that control of the symmetric
part of the gradient of u in L2 (Ω) is sufficient to control the entire gradient
of u in L2 (Ω).
Theorem 5.29. For any f ∈ L2 (Ω), there exists a unique weak solution
u ∈ H01 (Ω) to the homogeneous Dirichlet problem
∂ ij ∂u
− a =f in Ω , (5.23a)
∂xi ∂xj
u=0 on ∂Ω . (5.23b)
Moreover, the weak solution u satisfies
kukH 1 (Ω) ≤ Ckf kL2 (Ω) . (5.24)

Proof. According to Definition 5.8, we must find a unique element u ∈

H01 (Ω) satisfying
Z
B(u, v) = f (x)v(x) dx ∀ v ∈ H01 (Ω) ,
Ω
or simply Lu = f in H −1 (Ω) with L defined (5.18) (in which b ≡ 0). The

proof follows by an application of the Lax-Milgram Theorem; hence, we
must verify the coercivity and boundedness assumptions of that theorem.
Step 1. Boundedness. By Corollary 5.25, kDukL2 (Ω) is an equivalent
norm on H01 (Ω). Thus, we must find a constant C > 0 such that
kLukH −1 (Ω) ≤ CkDukL2 (Ω) ∀ v ∈ H01 (Ω) .
By the definition of the dual space norm,
∂u ∂v
Z
kLukH −1 (Ω) = sup hLu, viH01 (Ω) = sup aij (x) dx
kvkH 1 (Ω) =1 kvkH 1 (Ω) =1 Ω ∂xj ∂xi
0 0
≤ kakL∞ (Ω) kDukL2 (Ω) ,
where kakL∞ (Ω) ≡ max kaij kL∞ (Ω) .

i,j
∂u
Step 2. Coercivity. Let ξi = or equivalently, ξ = Du. By uniform
∂xi
ellipticity,
∂u ∂u
aij = aij ξj ξi ≥ λ 0 |ξ|2 = λ 0 |Du|2 .
∂xj ∂xi
Since, by definition, hLu, ui = B(u, u), we see that
∂u ∂u
Z
B(u, u) = aij dx ≥ λ 0 kDuk2L2 (Ω) = λ 0 kuk2H 1 (Ω) ∀ u ∈ H01 (Ω) .
Ω ∂xj ∂xi 0
In other words, L : H01 (Ω) → H −1 (Ω) is coercive if L is uniformly elliptic.

Note the essential use of the Poincaré inequality to show that B(u, u)
is coercive with respect to H01 (Ω). The fact that u = 0 on ∂Ω, which in
turn gives a trivial kernel for the operator L, is the key to the coercivity
inequality.
Theorem 5.30. For any f ∈ L2 (Ω), there exists a unique weak solution
u ∈ H 1 (Ω)/R to the homogeneous Neumann problem
∂ ij ∂u
− a =f in Ω , (5.25a)
∂xi ∂xj
∂u
aij Ni = 0 on ∂Ω . (5.25b)
∂xj
kukH 1 (Ω) ≤ Ckf kL2 (Ω) . (5.26)

H 1 (Ω)/R satisfying
Z
B(u, v) = f (x)v(x) dx ∀ v ∈ H 1 (Ω)/R , (5.27)
Ω
or simply Lu = f in the dual space of H 1 (Ω)/R with L defined (5.18) (in

which b ≡ 0). Again we verify the coercivity and boundedness assumptions
for L.
Step 1. (Boundedness) We must find a constant C > 0 such that
kLukH −1 (Ω) ≤ CkukH 1 (Ω) ∀ u ∈ H 1 (Ω)/R ,
while the proof of the inequality above is the same as the case of Dirichlet
problem.
Step 2. (Coercivity) By Theorem 5.22, kDukL2 (Ω) is an equivalent norm
on H 1 (Ω)/R. Thus, by uniform ellipticity,
∂u ∂u
Z
B(u, u) = aij dx ≥ λ 0 kDuk2L2 (Ω) ∀ u ∈ H 1 (Ω)/R .
Ω ∂xj ∂xi
In other words, L : H 1 (Ω)/R → (H 1 (Ω)/R)0 is coercive if L is uniformly

elliptic.
Remark 5.31. The use of the subspace H 1 (Ω)/R allows us to use the
Poincaré inequality and hence use kDukL2 (Ω) and an H 1 (Ω)-equivalent norm
on elements of H 1 (Ω)/R, and remove the one-dimensional kernel of the op-
erator L. Moreover, we can permit any forcing function f ∈ L2 (Ω). To see
that this is indeed the case, note that both f and f − c, for a constant c,
induce the same linear functional on H 1 (Ω)/R. In Section 5.6, we will show
that the weak formulation (5.27) is equivalent to
∂u ∂v
Z Z
ij
a dx = (f − c)v dx ∀ v ∈ H 1 (Ω) , (5.28)
Ω ∂xj ∂xi Ω
Z
where c = f¯ = − f (x) dx. Using (5.28) as the variational formulation of
Ω
the homogeneous Neumann problem, the forcing function F = f −c is indeed
required to satisfy the solvability condition to guarantee the invertibility of
L.
Theorem 5.32. If k ∈ L∞ (∂Ω) is non-negative on ∂Ω and is positive on a

portion of ∂Ω with positive measure, then for any f ∈ L2 (Ω), there exists a
unique weak solution u ∈ H 1 (Ω) to the Robin problem
∂ ij ∂u
− a =f in Ω, (5.29a)
∂xi ∂xj
∂u
aij Ni + ku = 0 on ∂Ω . (5.29b)
∂xj
kukH 1 (Ω) ≤ Ckf kL2 (Ω) . (5.30)

H 1 (Ω) satisfying
Z
B(u, v) + b(u, v) = f (x)v(x) dx ∀ v ∈ H 1 (Ω) ,
Ω
Z
where b is defined by b(u, ϕ) = kuϕ dS. This is the same as finding a
∂Ω
solution to Lu = f in H 1 (Ω)0 with L defined by (5.18).
Step 1. (Boundedness of L) We must find a constant C > 0 such that
kLukH −1 (Ω) ≤ CkukH 1 (Ω) ∀ u ∈ H 1 (Ω) .

By the Trace Theorem, a function w ∈ H 1 (Ω) belongs to L2 (∂Ω) and satis-

fies kwkL2 (∂Ω) ≤ kwkH 1 (Ω) . As a consequence,
kLukH 1 (Ω)0 = sup hLu, ϕiH 1 (Ω)

kϕkH 1 (Ω) =1
≤ kakL∞ (Ω) kDukL2 (Ω) + sup kkkL∞ (∂Ω) kukL2 (∂Ω) kϕkL2 (∂Ω)
kϕkH 1 (Ω) =1

≤ kakL∞ (Ω) + CkkkL∞ (∂Ω) kukH 1 (Ω) .
Step 2. (Coercivity of L) By uniform ellipticity,

√
B(u, u) + b(u, u) ≥ λ 0 kDuk2L2 (Ω) + k kuk2L2 (∂Ω) ∀ u ∈ H 1 (Ω) ,
so the coercivity of L amounts to the equivalency of the H 1 -norm and the

norm defined by
√
kwk ≡ k kwk2L2 (∂Ω) + kDwk2L2 (Ω) .
The equivalence of norms is guaranteed by the Poincaré inequality Lemma

5.24. Therefore, L : H 1 (Ω) → H 1 (Ω)0 is coercive if L is uniformly elliptic.
5.5.5 The non-standard elliptic problem
Theorem 5.33. Let ∂Ω be of class C 1 . Then for any f ∈ L2 (Ω), there

exists a unique weak solution u ∈ H 1 (Ω)/R ∩ H 1 (∂Ω) to the non-standard
elliptic problem
∂ ij ∂u
Lu = − a =f in Ω, (5.31a)
∂xi ∂xj
∂u
aij Ni = ∆g u on ∂Ω . (5.31b)
∂xj
kukH 1 (Ω) + kukH 1 (∂Ω) ≤ Ckf kL2 (Ω) . (5.32)

Proof. Let V = H 1 (Ω)/R ∩ H 1 (∂Ω) ≡ w ∈ H 1 (Ω)/R w ∈ H 1 (∂Ω) .

According to Definition 5.10, we must find a unique element u ∈ V satisfying

Z
B(u, v) + b(u, v) = f (x)v(x) dx ∀v ∈ V ,
Ω
Z
where b is defined by b(u, ϕ) = g(gradu, gradv)dS. This is the same as
∂Ω
finding a solution to Lu = f in V 0 with L defined (5.18).
h i
kLukV 0 ≤ C kukH 1 (Ω) + kukH 1 (∂Ω) ∀u ∈ V .
Since ∂Ω is of class C 1 , the metric tensor g ∈ L∞ (∂Ω). With kgkL∞ (∂Ω)

denoting max kgαβ kL∞ (∂Ω) , the Trace Theorem implies that
α,β

hLu, viV ≤ kakL∞ (Ω) kukH 1 (Ω) kϕkH 1 (Ω) + kgkL∞ (∂Ω) kukH 1 (∂Ω) kϕkH 1 (∂Ω)
h i
≤ kakL∞ (Ω) + kgkL∞ (∂Ω) kukV kϕkV ;
thus kLukV 0 ≤ CkukV for all u ∈ V.

Step 2. (Coercivity of L) By uniform ellipticity and the Poincaré in-
equality, for all u ∈ V,
1
hLu, uiV ≥ λ 0 kDuk2L2 (Ω) + kgraduk2L2 (∂Ω) ≥ kuk2H 1 (Ω) + kgraduk2L2 (∂Ω)
C1
for some constant C1 > 0. Moreover, the Trace Theorem implies that for
some constant C2 > 0, kukL2 (∂Ω) ≤ C2 kukH 1 (Ω) ; thus the inequality above
further implies that for some constant C > 0,
1h i
hLu, uiV ≥ kuk2H 1 (Ω) + kuk2H 1 (∂Ω) ∀u ∈ V
C
which guarantees the coercivity of L.
5.5.6 The elasticity equations

Dirichlet boundary conditions
Theorem 5.34. If the elasticity tensor C ijk` ∈ L∞ (Ω) satisfies the positivity
and symmetry assumptions of Definition 5.3, then for any f ∈ L2 (Ω; R3 ),
there exists a unique weak solution u ∈ H01 (Ω; R3 ) to the linear elasticity
equation
∂ ijk` ∂u`
− C = fi in Ω, (5.33a)
∂xj ∂xk
u=0 on ∂Ω . (5.33b)
kukH 1 (Ω;R3 ) ≤ Ckf kL2 (Ω;R3 ) .


H01 (Ω; R3 ) satisfying
∂u` ∂v i
Z Z
C ijk` dx = f (x) · v(x) dx ∀ v ∈ H01 (Ω; R3 ) .
Ω ∂xk ∂xj Ω
This is the same as finding a solution to Lu = f in H −1 (Ω; R3 ) if L is defined

by
∂u` ∂v i
Z
hLu, viH01 (Ω;R3 ) = C ijk` dx ∀ u, v ∈ H01 (Ω; R3 ) .
Ω ∂xk ∂xj
As before, we verify the coercivity and boundedness assumptions of L.
kLukH −1 (Ω;R3 ) ≤ CkukH 1 (Ω;R3 ) ∀ u ∈ H 1 (Ω; R3 ) .
Nevertheless, it follows from C ijk` ∈ L∞ (Ω) that
kLukH −1 (Ω;R3 ) ≤ max kC ijk` kL∞ (Ω) kDukL2 (Ω;R3 ) ∀ u ∈ H01 (Ω; R3 ) .
i,j,k,`
Step 2. (Coercivity of L) We note that the symmetry of C ijk` implies

that
C ijk` ui ,j u` ,k = C ijk` ij k` ;
thus the positivity of C ijk` and Korn’s first inequality implies that
1
hLu, uiH01 (Ω;R3 ) ≥ λkk2L2 (Ω) ≥ kuk2H 1 (Ω;R3 ) ∀ u ∈ H01 (Ω; R3 ) ;
C 0
so we establish the coercivity of L. Therefore, L : H01 (Ω; R3 ) → H −1 (Ω; R3 )

is coercive if the elasticity tensor C ijk` satisfies the positivity and the sym-
metry assumptions of Definition 5.3.
Neumann boundary conditions
there exists a unique weak solution u ∈ H 1 (Ω; R3 )/S to the linear elasticity
equation
∂ ijk` ∂u`
− C = fi in Ω, (5.34a)
∂xj ∂xk
∂u`
C ijk` Nj = 0 on ∂Ω . (5.34b)
∂xk
kukH 1 (Ω;R3 ) ≤ Ckf kL2 (Ω;R3 ) .
Proof. The proof for the case of the Neumann boundary condition is almost
the same as the Dirichlet case except that the coercivity of L is guaranteed
by Korn’s second inequality which implies that
1
hLu, uiH 1 (Ω;R3 ) ≥ kuk2H 1 (Ω;R3 ) ∀ u ∈ H 1 (Ω; R3 )/S.
C
5.6 Weak solutions with distributional forcing func-

tions
Recall that (·, ·)H denotes the inner product of the Hilbert space H, and
h·, ·iX denotes the duality pairing between X and its dual space X 0 .
When the interior forcing function f ∈ L2 (Ω) as in the previous section,
the operator L maps from a Hilbert space V into a proper subspace of the
dual space V 0 . Notice, however, that the Lax-Milgram Theorem allows us
to establish existence and uniqueness for elliptic BVPS when the forcing
function is given in the dual space V 0 of the solution space V.
In order to consider distributional forcing functions which are elements of
V 0 , it is necessary to explain the representation theory for these various dual
spaces. In the following example, we provide some representations of V 0 for
a variety of different boundary conditions that we have already discussed.
Example 5.36.
1. Let H1 = H01 (Ω), and f ∈ H −1 (Ω) = H01 (Ω)0 . By the Riesz represen-
tation theorem there exists a unique u ∈ H01 (Ω) such that
Z
hf, ϕiH1 = (Du, Dϕ)L2 (Ω) = Du · Dϕ dx ∀ ϕ ∈ H1 .
Ω
2. Let H2 = H 1 (Ω), and f ∈ H 1 (Ω)0 . By the Riesz representation theo-

rem there exists an u ∈ H 1 (Ω) such that
Z
hf, ϕiH2 = (u, ϕ)H 1 (Ω) = (uϕ + Du · Dϕ) dx ∀ ϕ ∈ H2 .
Ω
§5.6 Weak solutions with distributional forcing functions 207
n Z o
3. Let H3 = H (Ω)/R ≡ w ∈ H 1 (Ω)
1
w(x) dx = 0 be endowed with

Ω
the usual H 1 -norm, that is, kwkH3 = kwkH 1 (Ω) . Let f ∈ H30 , then by
the Riesz representation theorem, there exists an u ∈ H 1 (Ω)/R such
that
Z

hf, ψiH3 = (u, ψ)L2 (Ω) = uψ + Du · Dψ dx ∀ ψ ∈ H3 .
Ω
(5.35)
On the other hand, by the Poincaré inequality, (Du, Dψ)L2 (Ω) is an
equivalent inner product on H3 . Using this as the inner product on
H3 , for any f ∈ H30 , there exists v ∈ H 1 (Ω)/R such that
hf, ψiH3 = (Dv, Dψ)L2 (Ω) ∀ ψ ∈ H3 . (5.36)
4. Let f ∈ H 1 (Ω)0 , then f ∈ H30 ; thus there exists u ∈ H3 such that
hf, ψiH 1 (Ω) = (u, ψ)L2 (Ω) + (Du, Dψ)L2 (Ω) ∀ ψ ∈ H 1 (Ω)/R ,
where hf, ψiH 1 (Ω)/R = hf, ψiH 1 (Ω) is used to obtain the second equality.
In particular, for any ϕ ∈ H 1 (Ω), ψ ≡ ϕ − ϕ̄ ∈ H 1 (Ω)/R can be used
in the equality above, thus by (u, 1)L2 (Ω) = 0,
∀ ϕ ∈ H 1 (Ω) .

f, ϕ − ϕ̄ H 1 (Ω) = (u, ϕ)L2 (Ω) + (Du, Dϕ)L2 (Ω)
Let f ∈ C ∞ (Ω) and f → f in H 1 (Ω)0 . Then

Fubini’s Theorem
lim (f¯ , ϕ)L2 (Ω)

f, ϕ̄ H 1 (Ω)
= lim f , ϕ̄ L2 (Ω) =
→0 →0

= hf, 1iH 1 (Ω) , ϕ L2 (Ω) ;
thus u satisfies
f − hf, 1iH 1 (Ω) , ϕ H 1 (Ω) = (u, ϕ)L2 (Ω) + (Du, Dϕ)L2 (Ω) ∀ ϕ ∈ H 1 (Ω) ,

or if c = hf, 1iH 1 (Ω) ,
∀ ϕ ∈ H 1 (Ω) .

hf, ϕiH 1 (Ω) = (u + c, ϕ)L2 (Ω) + D(u + c), Dϕ L2 (Ω)
In other words, u + hf, 1iH 1 (Ω) is the representation of f ∈ H 1 (Ω)0 .

Moreover, even though in general f and f − hf, 1iH 1 (Ω) are different
bounded linear functionals on H 1 (Ω), they are the same in the dual
space of H 1 (Ω)/R.
5. Let H4 = H01 (Ω; Rn ), that is, u ∈ H4 if and only if u = (u1 , · · · , un )

and ui ∈ H01 (Ω) for i = 1, · · · , n. Then H4 is a Hilbert space endowed
n n
(ui , v i )H1 = (Dui , Dv i )L2 (Ω) .
P P
with the inner product (u, v)H4 =
i=1 i=1
If f = (f 1 , · · · , f n ) ∈ H −1 (Ω; Rn ) ≡ H01 (Ω; Rn )0 , by the Riesz repre-
sentation theorem, there exists a u = (u1 , · · · , un ) ∈ H01 (Ω; Rn ) such
that
n
X
hf, ϕiH4 = (u, ϕ)H4 = (Dui , Dϕ i )L2 (Ω) ∀ ϕ ∈ H4 .
i=1
1 (Ω; Rn ), that is, u ∈ H if and only if

6. Let H5 = Hdiv 5
(a) u ∈ Rn , u = (u1 , · · · , un ), ui ∈ H 1 (Ω) for i = 1, · · · , n,

n
ui ,i = 0 in Ω, and
P
(b) divu ≡
i=1
(c) u · N = 0 on ∂Ω.
H5 is a Hilbert space endowed with the inner product

n
X
(u, ϕ)H5 = (ui , ϕ i )H 1 (Ω)
i=1
n
X
(ui , ϕ i )L2 (Ω) + (Dui , Dϕ i )L2 (Ω)

= ∀ ϕ ∈ H5 .
i=1
1 (Ω; Rn )0 , by the Riesz representation theo-

If f = (f 1 , · · · , f n ) ∈ Hdiv
1 (Ω; Rn ) such that
rem, there exists a u = (u1 , · · · , un ) ∈ Hdiv
n
X
(ui , ϕ i )L2 (Ω) + (Dui , Dϕ i )L2 (Ω)

hf, ϕiH5 = ∀ ϕ ∈ H5 .
i=1
Example 5.36 gives some of the representation formulas for distributional

forcing functions. In the case that n = 1, a simple example is provided by
f =δ
where δ denotes the Dirac delta distribution. For example, with n = 1,

H01 (Ω) ⊆ C 0 (Ω) with continuous embedding, and hence C 0 (Ω)0 ⊆ H −1 (Ω),
so δ ∈ H −1 (Ω).
On the other hand, when f is a function in L2 (Ω), the above represen-
tations are not needed, as we discuss in the following remark.
§5.6 Weak solutions with distributional forcing functions 209
Remark 5.37. Let Hi be the Hilbert spaces in Example 5.36, and suppose
that a square integrable function f is an element in Hi0 . It follows that the
action of f on ϕ ∈ Hi is defined by hf, ϕiHi = (f, ϕ)L2 (Ω) . In particular a
function f ∈ L2 (Ω) defines a distribution Tf ∈ Hi0 by the formula
hTf , ϕiHi = (f, ϕ)L2 (Ω) .
With a slight abuse of notation, we will represent both functions f ∈ L2 (Ω)

and the induced distribution Tf by the symbol f .
With the notion of distributional forcing, we have the following theorems

which generalize Theorem 5.29, 5.30, 5.32, 5.33, 5.34 and 5.35.
Theorem 5.38. For any f ∈ H −1 (Ω), there exists a unique weak solution
u ∈ H01 (Ω) to the homogeneous Dirichlet problem (5.23) satisfying
B(u, v) = hf, viH01 (Ω) ∀ v ∈ H01 (Ω) ,
kukH 1 (Ω) ≤ Ckf kH −1 (Ω) . (5.37)
Theorem 5.39. For any f in the dual space of H 1 (Ω)/R, there exists a
unique solution u ∈ H 1 (Ω)/R to
B(u, v) = hf, viH 1 (Ω)/R ∀ v ∈ H 1 (Ω)/R ,
kukH 1 (Ω) ≤ Ckf k(H 1 (Ω)/R)0 . (5.38)
In particular, for any f ∈ H 1 (Ω)0 satisfying hf, 1iH 1 (Ω) = 0, there exists a
unique solution u ∈ H 1 (Ω)/R to
B(u, v) = hf, viH 1 (Ω)0 ∀ v ∈ H 1 (Ω)
satisfying kukH 1 (Ω) ≤ Ckf kH 1 (Ω)0 .

Theorem 5.40. If k ∈ L∞ (∂Ω) is non-negative on ∂Ω and is positive on

a portion of ∂Ω, then for any f ∈ H 1 (Ω)0 , there exists a unique solution
u ∈ H 1 (Ω) to
Z
B(u, v) + kuvdS = hf, viH 1 (Ω) ∀ v ∈ H 1 (Ω) .
∂Ω
kukH 1 (Ω) ≤ Ckf kH 1 (Ω)0 . (5.39)
Theorem 5.41. Let ∂Ω be of class C 1 . Then for any f in the dual space of
H 1 (Ω)/R ∩ H 1 (∂Ω), there exists a unique solution u ∈ H 1 (Ω)/R ∩ H 1 (∂Ω)
to
Z
B(u, v) + g(gradu, gradv) dS = hf , viH 1 (Ω)/R∩H 1 (∂Ω)
∂Ω
∀ v ∈ H 1 (Ω)/R ∩ H 1 (∂Ω) .
kukH 1 (Ω) + kukH 1 (∂Ω) ≤ Ckf k(H 1 (Ω)/R∩H 1 (∂Ω))0 . (5.40)
and symmetry assumptions of Definition 5.3, then for any f ∈ H −1 (Ω; R3 ),
there exists a unique weak solution u ∈ H01 (Ω; R3 ) to the linear elasticity
equation (5.33) satisfying
Z
C ijk` u` ,k v i ,j dx = hf, viH01 (Ω;R3 ) ∀ v ∈ H01 (Ω; R3 ) .
Ω
kukH 1 (Ω;R3 ) ≤ Ckf kH −1 (Ω;R3 ) .
there exists a unique solution u ∈ H 1 (Ω; R3 )/S to
Z
C ijk` u` ,k v i ,j dx = hf, viH 1 (Ω;R3 )/S ∀ v ∈ H01 (Ω; R3 )/S .
Ω
kukH 1 (Ω;R3 ) ≤ Ckf k(H 1 (Ω;R3 )/S)0 .

5.7. GENERAL SECOND ORDER ELLIPTIC OPERATORS 211
5.7 General second order elliptic operators
Suppose now we want to solve

∂ ij ∂u ∂u
Lu = − a + bi + cu = f in Ω, (5.41a)
∂xi ∂xj ∂xi
u=0 on ∂Ω , (5.41b)
where L is strictly elliptic, a, b, c ∈ L∞ (Ω). Let L : H01 (Ω) → H −1 (Ω) be

defined by for ϕ ∈ H01 (Ω),
ij ∂v ∂ϕ i ∂v
Z Z Z
hLv, ϕiH01 (Ω) = a dx + b ϕdx + cvϕ dx .
Ω ∂xj ∂xi Ω ∂xi Ω
It is easy to see that L : H01 (Ω) → H −1 (Ω) is continuous. In fact,
kLukH −1 (Ω) = sup hLu, ϕiH01 (Ω)

kϕkH 1 (Ω) =1
0
h i
≤ max kaij kL∞ (Ω) + kbkL∞ (Ω) + kckL∞ (Ω) kukH01 (Ω) .
i,j
However, it is not clear that if there exists γ0 > 0 so that γkuk2H 1 (Ω) ≤
0
hLu, uiH01 (Ω) . In fact, if c 0, L fails to be coercive. Let I : H01 (Ω) →
H −1 (Ω) be defined by
Z
hIv, ϕiH01 (Ω) = vϕ dx .
Ω
Then we have the following
Lemma 5.44. For some γ > 0 sufficient large, (L + γI) : H01 (Ω) → H −1 (Ω)
is coercive.
Proof. By the positivity of aij ,

∂u ∂u
Z
λ 0 kDuk2L2 (Ω) ≤ aij dx
Ω ∂xj ∂xi
Z Z
= hLu, uiH01 (Ω) − (b · Du)udx − cu2 dx
Ω Ω
≤ hLu, uiH01 (Ω) + kbkL∞ (Ω) kDukL2 (Ω) kukL2 (Ω) + kckL∞ (Ω) kuk2L2 (Ω)
λ0 h 1 i
≤ hLu, uiH01 (Ω) + kuk2H 1 (Ω) + kbk2L∞ (Ω) + kckL∞ (Ω) kuk2L2 (Ω) ,
2 0 2λ 0
where we use Young’s inequality
1 2
ab ≤ a2 + b ∀ > 0
4
1
to obtain the last inequality. Choose γ > kbk2L∞ (Ω) + kckL∞ (Ω) to con-
2λ 0
clude that
λ0
kuk2H 1 (Ω) ≤ hLu, uiH01 (Ω) + γkuk2L2 (Ω) = hLu + γIu, uiH01 (Ω) .
2 0
Theorem 5.45. For all f ∈ H −1 (Ω), there exists unique solution u ∈

H01 (Ω) to (L + γI)u = f in H −1 (Ω) if γ is large enough (independent of f ).
From now on, the solution u ∈ H01 (Ω) to (L + γI)u = f in H −1 (Ω)

is denoted by (L + γI)−1 f . The goal for the remaining section is to find
conditions of f ∈ L2 (Ω) under which Lu = f have a unique solution.
Definition 5.46. We call L∗ the formal adjoint of L and define it by
(L∗ v, u)L2 (Ω) = (Lu, v)L2 (Ω) ∀ u, v ∈ H01 (Ω) .
Integrating by parts, since u, v vanish on the boundary,
ij ∂u ∂v i ∂u
Z Z Z
(Lu, v)L2 (Ω) = a dx + b vdx + cuvdx
∂ ij ∂v ∂
Z Z Z
i
= − a udx + u (−b v)dx + cuvdx .
Therefore,
∂ ij ∂v ∂ i
L∗ v = − a − (b v) + cv .
∂xj ∂xi ∂xi
Theorem 5.47 (Fredholm Alternative). L2 (Ω) = N (L∗ ) ⊕ R(L), where

n o
N (L∗ ) = v ∈ H01 (Ω) L∗ v = 0 .

So either there exists a unique solution to Lu = f ∈ L2 (Ω) (the case

N (L∗ ) = {0}) or there is a v 6= 0 such that L∗ v = 0.
Theorem 5.47 follows from the theory of compact operator: for compact
operator K : H → H, H Hilbert (see Appendix B.3 for details),
(1) K ∗ is also compact;
(2) dim N (I − K) = dim(N (I − K ∗ )) < ∞;
(3) R(I − K) = R(Id − K ∗ ) is closed;

§5.7 General second order elliptic operators 213
(4) N (I − K ∗ )⊥ = R(I − K);
(5) (I − K)u = h has a unique solution in H or ∃ v 6= 0 in H solving

(I − K ∗ )v = 0.
Theorem 5.48 (Fredholm Alternative: another version). For all f ∈ L2 (Ω),

there exists a unique solution u ∈ H01 (Ω) to Lu = f or N (L∗ ) is not trivial.
Proof. Lu = f has a weak solution if and only if Lu + γu = f + γu has a

weak solution. By Theorem 5.45, we may choose γ > 0 so that (L + γI)−1
exists and
u = (L + γI)−1 f + γ(L + γI)−1 u or [I − γ(L + γI)−1 ]u = (L + γI)−1 f .
Claim: γ(L + γI)−1 : L2 (Ω) → L2 (Ω) is compact.
Again by Theorem 5.45, for all g ∈ L2 (Ω), there exists a unique solution u
to (L + γI)u = g and
kukH01 (Ω) = k(L + γI)−1 gkH01 (Ω) ≤ CkgkL2 (Ω) .
So (L + γI)−1 is a bounded linear operator from L2 (Ω) → H01 (Ω). By

Rellich’s theorem, H01 (Ω) is compactly embedded into L2 (Ω), so γ(L+γI)−1 :
L2 (Ω) → L2 (Ω) is compact. Therefore, (5) implies that it suffices to show
that
∃ v 6= 0 3 I − γ(L + γI)−1∗ v = 0 ⇔ {0} ( N (L∗ ) .

However, I − γ(L + γI)−1∗ v = 0 ⇔ (L + γI)∗ v = γv ⇔ L∗ v = 0.

Remark 5.49.
1. By fact (2), the last line of the proof also shows that dim(N (L∗ )) =
dim(N (L)) < ∞.
2. By fact (4), Lu = f has a solution (or equivalently, u solves (I −

γ(L + γI)−1 )u = (L + γI)−1 f ) if (L + γI)−1 f is orthogonal to v for all
v ∈ N (I − γ(L + γI)−1∗ ), i.e., ((L + γI)−1 f, v)L2 (Ω) = 0 for all such v.
But
0 = (L + γI)−1 f, v = f, (L + γI)−1∗ v

L2 (Ω) L2 (Ω)
= γ(f, v)L2 (Ω) .
What we have established is Fredholm Alternative theorem L2 (Ω) =

R(L) ⊕ N (L∗ ).
3. Recall we have proved this for open bounded Ω. For unbounded Ω, the
Fredholm Alternative is not necessary true due to failure of Rellich’s
theorem.
4. In the proof of the Fredholm Alternative, the use of Dom(L) = H01 (Ω)
is not essential: it can be replaced by other spaces (but it might be dif-
ficult to compute the formal adjoint L∗ ). The same proof goes through
if Rellich’s theorem (or compactness arguments) can still be applied.
Theorem 5.50. Suppose that f ∈ L2 (Ω).
(1) If N (L∗ ) = {0}, then there exists a unique weak solution u ∈ H01 (Ω)
to Lu = f in H −1 (Ω). Furthermore,
kukH 1 (Ω) ≤ Ckf kL2 (Ω) . (5.42)
(2) If N (L∗ ) 6= {0} and f ⊥ N (L∗ ), then there exist multiple weak solu-
tions in H01 (Ω) to Lu = f , and each solution u satisfies
h i
kukH 1 (Ω) ≤ C kukL2 (Ω) + kf kL2 (Ω) . (5.43)
(3) If N (L∗ ) 6= {0} and f 6⊥ N (L∗ ), then there exists no weak solution in
H01 (Ω) to Lu = f .
Proof. (3) is proved in the previous Remark. Now suppose f ⊥ N (L∗ ) and
u is a weak solution to Lu = f . Then
Z Z Z Z
aij u,j v,i dx + bi u,i v dx + cuv dx = f v dx .
Ω Ω Ω Ω
Let v = u. Then
λ 0 kDuk2L2 (Ω)
1
≤ kbkL∞ (Ω) kDukL2 (Ω) kukL2 (Ω) + kckL∞ (Ω) kuk2L2 (Ω) + kf k2L2 (Ω) + kuk2L2 (Ω)

2
λ0 h 1 1 i 1
≤ kDuk2L2 (Ω) + kbk2L∞ (Ω) + kckL∞ (Ω) + kuk2L2 (Ω) + kf k2L2 (Ω) .
2 2λ 0 2 2
Therefore, by Poincaré’s inequality,
h i
kuk2H 1 (Ω) ≤ CkDuk2L2 (Ω) ≤ C kuk2L2 (Ω) + kf k2L2 (Ω) .
5.8. HORIZONTAL CONVOLUTION-BY-LAYERS AND COMMUTATION ESTIMATES215
This proves (2).

Now suppose that N (L∗ ) = {0}. Then fact (4) suggests that R(I −
γ(γL + γI)−1 ) = L2 (Ω); hence by the bounded inverse theorem B.14, the
inverse of I − γ(γL + γI)−1 belongs to B(L2 (Ω), L2 (Ω)). Since u = I −

−1
γ(L + γI)−1 (L + γI)−1 f ,
kukL2 (Ω) ≤ Ck(L + γI)−1 f kL2 (Ω) ≤ Ckf kL2 (Ω) .
Estimate (5.42) then follows from (5.43) and the L2 -estimate above.
5.8 Horizontal convolution-by-layers and commu-

tation estimates
5.8.1 Horizontal convolution-by-layers

With xh = (x1 , · · · , xn−1 ), we define ρ ∈ C0∞ (Rn−1 ) by ρ(xh ) = C exp 1
|xh |2 −1
ifZ |xh | < 1 and ρ(xh ) = 0 if |xh | ≥ 1; we then select the constant C so that
ρ dxh = 1. We define ρ (xh ) = −(n−1) ρ(−1 xh ). It follows that for
Rn−1
> 0, 0 ≤ ρ ∈ C0∞ (R2 ) with spt(ρ ) ⊆ B(0, ).
We define the horizontal convolution-by-layers operator Λ as follows:
Z
Λ f (xh , xn ) = ρ (xh − yh )f (yh , xn )dyh for f (·, xn ) ∈ L1 (Rn−1 ) .
Rn−1
It should be clear that Λ smooths functions defined on Rn along all hor-

izontal subspaces, but does not smooth these function in the vertical xn -
direction. On the other hand, we can restrict the operator Λ to act on
functions f : Rn−1 → R as well, in which case Λ becomes the usual mollifi-
cation operator.
By standard properties of convolution, there exists a constant C which
is independent of , such that for s ≥ 0,
kΛ F kH s (Ω) ≤ CkF kH s (Ω) ∀ F ∈ H s (Ω) ,
and
kΛ F kH s (∂Ω) ≤ CkF kH s (∂Ω) ∀ F ∈ H s (∂Ω) .
Furthermore,
¯ F kL2 (Ω) ≤ CkF kL2 (Ω)

k∂Λ ∀ F ∈ L2 (Ω) , (5.44)
where ∂¯ denotes the vector of horizontal partial derivatives:

¯ ∂ ∂
∂= , · · ·, .
∂x1 ∂xn−1
5.8.2 Commutation estimates
Definition 5.51. Let X be a Banach space, and A : X → X and B : X → X

be two operators. The commutator of A and B, denoted by A, B , is defined
by

A, B w = A(Bw) − B(Aw) ∀w ∈ X . (5.45)
Lemma 5.52. For f ∈ W 1,∞ (Rn+ ) and g ∈ L2 (Rn+ ), there is a generic

constant C independent of such that
∂¯ Λ , f g 2 n = ∂¯ Λ (f g) − f Λ g 2 n

L (R+ ) L (R+ )

≤ C f 1,∞ n g 2 n
W (R+ ) L (R+ )
(5.46)
for all > 0.
Proof. By definition,
Z

Λ , f g(x) = ρ (xh − yh ) f (yh , xn ) − f (xh , xn ) g(yh , xn ) dyh .
B(xh ,)

By Morrey’s inequality, f (yh , xn ) − f (xh , xn ) ≤ Ckf kW 1,∞ (Rn+ ) . Since
¯ (xh − yh ) = 1 (Dρ) xh − yh ,

∂ρ
n
we see that
Z
1 x − y
∂¯ Λ , f g (x) ≤ Ckf kW 1,∞ (Rn )
h h
(Dρ) g(yh , xn )dyh .
+
Rn−1 n−1
Finally, we note that
Z
1 x − y Z
h h
n−1
|Dρ| dyh = |Dρ|(zh )dzh = kDρkL1 (Rn−1 ) ;
Rn−1 Rn−1
thus by Young’s inequality,
∂¯ Λ , f g (·, xn ) 2 n−1

L (R )
1 x − ·
h
≤ Ckf kW 1,∞ (Rn+ ) n−1 (Dρ) 1 n−1 kg(·, xn )kL2 (Rn−1 )

L (R )
≤ Ckf kW 1,∞ (Rn+ ) kg(·, xn )kL2 (Rn−1 ) .
5.9. ELLIPTIC REGULARITY 217
The desired estimate (5.46) then follows from integrating the square of the
inequality above in xn over the interval (0, ∞).
The same proof leads to the following
Lemma 5.53. For f ∈ W 1,∞ (Ω) and g ∈ L2 (Ω) with compact support,
there is a generic constant C independent of such that

D η ∗, f g 2 = D η ∗ (f g) − f η ∗ g 2
L (Ω) L (Ω)
≤ Ckf kW 1,∞ (Ω) kgkL2 (Ω) (5.47)

for all < min dist ∂Ω, spt(f ) , dist ∂Ω, spt(g) .
5.9 Elliptic regularity

5.9.1 Dirichlet problem with homogeneous boundary condi-
tion
Theorem 5.54. Suppose that Ω ⊆ Rn is an open, bounded domain with

smooth boundary, L is uniformly elliptic, and aij ∈ C 1 (Ω). Then for any
f ∈ L2 (Ω), the unique weak solution u to the Dirichlet problem (5.23) in
fact belongs to H 2 (Ω) ∩ H01 (Ω) and satisfies
kukH 2 (Ω) ≤ Ckf kL2 (Ω) , (5.48)
where C depends on kaij kC k (Ω) and the regularity of ∂Ω.
Proof. The goal is to show that the function u ∈ H01 (Ω) satisfying
Z Z
ij
a u,j v,i dx = f vdx ∀ v ∈ H01 (Ω) (5.49)
Ω Ω
satisfies (5.48). We divide the proof into several steps as follows:

Step 1. Interior estimates. We show that for any given Ω0 ⊂⊂Ω, the weak
solution u ∈ H 2 (Ω0 ) and satisfies
kukH 2 (Ω0 ) ≤ Ckf kL2 (Ω) , (5.50)
where C depends on Ω0 and in general C → ∞ as dist(∂Ω, Ω0 ) → 0. Interior

estimates are performed away from the boundary using cut-off functions
with compact support. Consequently, we are permitted to use integration

by parts in all directions to obtain such estimates.
Step 2. Estimates near the boundary. Estimates in regions near the bound-
ary are significantly more delicate. In such regions, we remap the local neigh-
borhood of the boundary into the upper-half of the unit sphere, and perform
integration by parts only in horizontal directions – the normal derivatives
are then estimated using the structure of the elliptic operator, and in small
open neighborhood U of ∂Ω, we find that
kukH 2 (U ) ≤ Ckf kL2 (Ω) (5.51)
The combination of (5.50) and (5.51) then leads to (5.48).

Step 1. Interior estimates.
Choose any (nonempty) open sets Ω1 ⊂⊂Ω2 ⊂⊂Ω and let ζ ∈ C0∞ (Ω2 )
with 0 ≤ ζ ≤ 1 and ζ = 1 on Ω1 . Let 0 = min dist(spt(ζ), ∂Ω2 )/2. For all
0 < < 0 , define u (x) = η ∗ u(x) for all x ∈ Ω2 , and set
v = η ∗ (ζ 2 u ,k ),k . (5.52)
Then v ∈ H01 (Ω) and can be used as a test function in (5.49); thus
Z Z
ij 2
f η ∗ (ζ 2 u ,k ),k dx

a u,j η ∗ (ζ u ,k ),ki dx =
Ω Ω
or
Z Z
aij u,j η ∗(ζ 2 u ,i ),kk dx + aij u,j η ∗ (2ζζ,i u ),kk dx

Ω Z Ω Z
ij
f η ∗ (ζ 2 u ,k ),k dx .

+ a u,j η ∗ (2ζζ,k u ),ki dx =
Ω Ω
The property of the convolution allows us to rewrite the equality above as

Z Z
η ∗ (aij u,j ) ,k (ζ 2 u ,i ),k dx = − f η ∗ (ζ 2 u ,k ),k dx

Ω Z Ω Z
ij
aij u,j η ∗ (2ζζ,k u ),ki dx .

+ a u,j η ∗ (2ζζ,i u ),kk dx +
Ω Ω

Given the commutator ·, · defined in Definition 5.51,
η ∗ (aij u,j ) = η ∗, aij u,j + aij u ,j .

§5.9 Elliptic regularity 219
Therefore, we may further rewrite the previous integral equality as

Z Z
ζ a u ,jk u ,ik dx + 2 aij u ,jk ζζ,k u ,i dx
2 ij
Ω Z Ω Z
ij 2
η ∗, aij u,j ,k (ζ 2 u ,i ),k dx

+ a ,k u ,j (ζ u ,i ),k dx +
Ω Z Ω Z
2
aij u,j η ∗ (2ζζ,i u ),kk dx

= − f η ∗ (ζ u ,k ),k dx +
ΩZ Ω
aij u,j η ∗ (2ζζ,k u ),ki dx .

+
Ω
By the commutation estimate (5.47),
η ∗, aij u,j ,k 2

L (Ω)
≤ CkakC 1 (Ω) kukH 1 (Ω) .
Therefore, by uniform ellipticity, the Cauchy-Young inequality, and the H 1 -

estimate (5.24), we find that
h
λ 0 kζD2 u k2L2 (Ω) ≤ C max kaij kL∞ (Ω) kukH 1 (Ω) kζD2 u kL2 (Ω)
i,j
+ kakC 1 (Ω) kukH 1 (Ω) kζD2 ukL2 (Ω)

i
+ kf kL2 (Ω) kζD2 ukL2 (Ω) + kukH 1 (Ω)

λ0 C
kζD2 u k2L2 (Ω) + kak2C 1 (Ω) + 1 kf k2L2 (Ω) ,

≤
2 λ0
where we remark that C depends on the derivatives of ζ; thus C = C(Ω1 , Ω2 ).
Since ζ = 1 in Ω1 ,
C(Ω1 , Ω2 )
kD2 u kL2 (Ω1 ) ≤

kakC 1 (Ω) + 1 kf kL2 (Ω) .
λ0
Since the right-hand side does not depend on > 0, there exists a subse-
quence
0
D2 u * w in L2 (Ω) .
By Theorem 2.78, u → u in H 1 (Ω1 ), so that w = D2 u on Ω1 . As weak

convergence is lower semi-continuous, kD2 ukL2 (Ω1 ) ≤ Ckf kL2 (Ω) . As Ω1 and
2 (Ω). For any w ∈ H 1 (Ω),
Ω2 are arbitrary, we have established that u ∈ Hloc 0
2 (Ω), we may integrate by parts to find
set v = ζw in (5.49). Since u ∈ Hloc
that Z
aij u,j ,i +f ζw dx = 0 ∀ w ∈ H01 (Ω) .

Ω
Since w is arbitrary, and the spt(ζ) can be chosen arbitrarily close to ∂Ω, it
follows that for all x in the interior of Ω, we have that
∂ ij ∂u
− a (x) (x) = f (x) for almost every x ∈ Ω . (5.53)
∂xi ∂xj
Step 2. Boundary estimates.
We proceed to establish the regularity of u all the way to the boundary
∂Ω. Let {Um }K
m=1 denote an open cover of Ω which intersects the boundary
∂Ω, and let {θm }K
m=1 denote a collection of charts such that
1. θm : B(0, rm ) → Um is a C ∞ -diffeomorphism,
2. det Dθm = 1,
3. θm (B(0, rm ) ∩ {xn = 0}) → Um ∩ ∂Ω,
4. θm (B(0, rm ) ∩ {xn > 0}) → Um ∩ Ω.
Let 0 ≤ ζm ≤ 1 in C0∞ (Um ) denote a partition of unity subordinate to the

open covering Um .
Substitute the test function
−1
v = Λ [(ζm ◦ θm )2 Λ (u ◦ θm ),α ],α ◦ θm ∈ H01 (Ω)
into (5.49), and use the change of variables formula to obtain the identity
Z Z
ij ` k
a Aj (u◦θm ),` Ai (v ◦θm ),k dx = f ◦θm v ◦θm dx , (5.54)
B+ (0,rm ) B+ (0,rm )
where the C ∞ -matrix A(x) = [Dθm (x)]−1 and B+ (0, rm ) = B(0, rm )∩{xn >
0}. We define um = u ◦ θm , and bk` = aij A`j Aki . Then, with ξm = ζm ◦ θm ,
we can rewrite the test function as
2
Λ um ,α ,α .

v ◦ θm = Λ ξm
Since differentiation commutes with convolution, we have that
2
(v ◦ θm ),k = Λ (ξm Λ um ,kα ),α +2Λ (ξm ξm ,k Λ um ,α ),α ,
and we can express the left-hand side of (5.54) as

Z
bk` (u ◦ θm ),` (v ◦ θm ),k dx = I1 + I2 ,
B+ (0,rm )
where
Z
I1 = bk` um ,` Λ (ξm
2
Λ um ,kα ),α dx ,
B+ (0,rm )
Z
I2 = 2 bk` um ,` Λ (ξm ξm ,k Λ um ,α ),α dx .
B+ (0,rm )
Next, we see that

Z
I1 = [Λ (bk` um ,` )],α (ξm
2
Λ um ,kα ) dx = I1a + I1b ,
B+ (0,rm )
where
Z
I1a = (bk` Λ um ,` ),α ξm
2
Λ um ,kα dx ,
B+ (0,rm )
Z
Λ , bk` um ,` ,α ξm
2
Λ um ,kα dx ,

I1b =
B+ (0,rm )
and where Λ , bk` is the commutator of the horizontal convolution operator

and multiplication. The integral I1a produces the positive sign-definite term
which will allow us to build the global regularity of U , as well as an error
term:
Z h i
2 k`
I1a = ξm b Λ um ,`α Λ um ,kα + bk` ,α Λ um ,` ξm
2
Λ um ,kα dx ;
B+ (0,rm )
thus, together with the right hand-side of (5.54), we see that

Z Z

2 k` m m k` m 2 m
ξm b Λ u ,`α Λ u ,kα dx ≤ b ,α Λ u ,` ξm Λ u ,kα dx

B+ (0,rm ) B+ (0,rm )
Z

+ |I1b | + |I2 | + f ◦ θm v ◦ θm dx .

B+ (0,rm )
Since each θm is a C ∞ -diffeomorphism, it follows that the matrix bk` is

positive definite: there exists λ > 0 such that
λ|Y |2 ≤ bk` Yk Y` ∀Y ∈ Rn .
It follows that
Z Z

2 ¯
λ ξm |∂DΛ um |2 dx ≤ bk` ,α Λ um ,` ξm
2
Λ um ,kα dx

B+ (0,rm ) B+ (0,rm )
Z

+ |I1b | + |I2 | + f ◦ θm v ◦ θm dx ,

B+ (0,rm )
where D = (∂x1 , ..., ∂xn ) and ∂¯ = (∂x1 , ..., ∂xn−1 ). Application of the Cauchy-
Young inequality with δ > 0 shows that
Z Z

bk` ,α Λ um ,` ξm2
Λ um ,kα ] dx + |I2 | + f ◦ θm v ◦ θm dx

B+ (0,rm ) B+ (0,rm )
Z
2 ¯
≤δ ξm |∂DΛ um |2 dx + Cδ kf k2L2 (Ω) .
B+ (0,rm )
It remains to establish such an upper bound for |I1b |. Nevertheless, by the

commutator estimate (5.46),

∂xα Λ , bk` um ,` 2

L (B+ (0,rm ))
≤ CkbkW 1,∞ ((B+ (0,rm )) kukH 1 (Ω) ≤ CkakC 1 (Ω) kf kL2 (Ω) .
It then follows from the Cauchy-Young inequality with δ > 0 that

Z
2 ¯
|I1b | ≤ δ ξm |∂DΛ um |2 dx + Cδ kf |2L2 (Ω) .
B+ (0,rm )
By choosing 2δ < λ, we obtain the estimate

Z
2 ¯
ξm |∂DΛ um |2 dx ≤ Cδ kf k2L2 (Ω) .
B+ (0,rm )
Since the right hand-side is independent of , we find that

Z
2 ¯
ξm |∂Dum |2 dx ≤ Cδ kf k2L2 (Ω) . (5.55)
B+ (0,rm )
From (5.53), we know that − aij (x)u,j (x) ,i = f (x) for a.e. x ∈ Um .

By the chain-rule this means that almost everywhere in B+ (0, rm ),
−bk` um ,k` = bk` ,k um ,` +f ◦ θm ,
or equivalently,
−bnn um kα m β` m k` m
nn = b u ,αk +b u ,`β +b ,k u ,` +f ◦ θm .
If en = (0, · · · , 0, 1), then bnn = eT 2

n b en ≥ λ|en | = λ > 0; thus it follows
from (5.55) that
Z
2
ξm |D2 um |2 dx ≤ Cδ kf k2L2 (Ω) . (5.56)
B+ (0,rm )
Summing over m from 1 to K and combining with our interior estimates,

we have that
kukH 2 (Ω) ≤ Ckf kL2 (Ω) .

5.9.2 Dirichlet problem with inhomogeneous boundary con-

dition
Next we want to solve the following Dirichlet problem with inhomogeneous

boundary condition:
∂ ij ∂u
− a =f in Ω, (5.57a)
∂xi ∂xj
u=g on ∂Ω , (5.57b)
where f ∈ L2 (Ω) and g ∈ H 1.5 (∂Ω). By Theorem 2.156, there exists φ ∈

H 2 (Ω) so that φ = g on ∂Ω and kφkH 2 (Ω) ≤ CkgkH 1.5 (∂Ω) . Then w = u − φ
is the solution to
∂ ij ∂w ∂ ij ∂φ
− a =f+ a in Ω , (5.58a)
∂xi ∂xj ∂xi ∂xj
w=0 on ∂Ω , (5.58b)
hence u satisfies the estimate

h i h i
kuk2H 2 (Ω) ≤ C kwk2H 2 (Ω) + kφk2H 2 (Ω) ≤ C kf k2L2 (Ω) + kgk2H 1.5 (∂Ω) . (5.59)
5.9.3 Neumann problem
In this subsection we consider the Neumann problem
∂ ij ∂u
− a =f in Ω, (5.60a)
∂xi ∂xj
∂u
aij Ni = g on ∂Ω , (5.60b)
∂xj
Note that the weak formulation for the Neumann problem is
ij ∂u ∂ϕ
Z Z Z
a dx = f v dx + gv dS ∀ v ∈ H 1 (Ω) .
Ω ∂x j ∂xi Ω ∂Ω
In order to obtain the corresponding

Z regularity result, it suffices to estimate
the additional boundary integral gv dS for test functions
∂Ω
1. v = η ∗ (ζ 2 u ,k ),k in the proof of the interior estimates, or

−1
2. v = Λ [(ζm ◦ θm )2 Λ (u ◦ θm ),α ],α ◦ θm in the proof of the boundary
estimates.
There is nothing we have to do if v has compact support in Ω, so we

consider the case that
−1
v = Λ [(ζm ◦ θm )2 Λ (u ◦ θm ),α ],α ◦ θm
for some fixed m = 1, 2, · · · , K. However, it follows from H 0.5 (∂Ω)-H −0,5 (∂Ω)
duality the Trace Theorem, and the Cauchy-Young inequality that
Z
gϕ dS ≤ CkgkH 0.5 (∂Ω) ζm Λ um ,α H 0.5 (∂Ω∩Um )

∂Ω
≤ Cδ kgk2H 0.5 (∂Ω) + δ ζm ∂xα Λ um H 1 (Um )

h i 2
≤ Cδ kf k2L2 (Ω) + kgk2H 0.5 (∂Ω) + δ ζm ∂xα DΛ um L2 (Um ) .

By choosing δ > 0 small enough, we obtain

h i
kukH 2 (Ω) ≤ C kf kL2 (Ω) + kgkH 0.5 (∂Ω) . (5.61)
5.9.4 The presence of b and c in the elliptic operator L
The H 2 -regularity of the weak solution to the Dirichlet problem with bi =

c = 0 for all i is studied in Theorem 5.54. With the inclusion of b and c,
the existence of a solution is non-trivial; nevertheless, can be verified by
the Fredholm alternative. As long as the existence of a weak solution is
guaranteed, we are interested in the regularity of weak solutions. We have
the following
Theorem 5.55. Assume that
aij , bi , c ∈ C 1 (Ω) , f ∈ L2 (Ω) .
Suppose that u ∈ H01 (Ω) is a weak solution of the Dirichlet problem (5.41),
and ∂Ω is of class C 2 . Then u ∈ H 2 (Ω) ∩ H01 (Ω) and satisfies the estimate
h i
kukH 2 (Ω) ≤ C kf kL2 (Ω) + kukL2 (Ω) (5.62)
for some constant C depending only on m, U , and the coefficient of L.
Proof. If u is a weak solution to (5.41), then

Z Z
aij u,j ϕ,i dx = F ϕ dx ∀ ϕ ∈ H01 (Ω) ,
Ω Ω
where F = f − b · Du − cu . Note that this is the same as saying that u is

the unique weak solution to the Dirichlet problem
∂ ij ∂u
− a =F in Ω ,
∂xi ∂xj
u=0 on ∂Ω .
Therefore, by Theorem 5.54,
kukH 2 (Ω) ≤ CkF kL2 (Ω) ,
and (5.62) follows from Theorem 5.50 that

h i h i
kF kL2 (Ω) ≤ C kf kL2 (Ω) + kukH 1 (Ω) ≤ C kf kL2 (Ω) + kukL2 (Ω) .
5.9.5 Higher regularity
The proof to the following two theorems, depending on choices of different

test functions, are similar to the proof of Theorem 5.54 so we state without
proving the theorems.
Theorem 5.56. Let m be a non-negative integer, and assume that
aij , bi , c ∈ C m+1 (Ω) , f ∈ H m (Ω) .
Suppose that u ∈ H01 (Ω) is a weak solution of the Dirichlet problem (5.41),
and ∂Ω is of class C m+2 . Then u ∈ H m+2 (Ω) ∩ H01 (Ω) and satisfies the
estimate
h i
kukH m+2 (Ω) ≤ C kf kH m (Ω) + kukL2 (Ω)
for some constant C depending only on m, Ω, and the coefficient of L.
Theorem 5.57. Let m be a non-negative integer, and assume that
aij , bi , c ∈ C m+1 (Ω) , f ∈ H m (Ω) .
Suppose that u ∈ H 1 (Ω) is a weak solution of the boundary value problem

∂ ij ∂u ∂u
− a + bi + cu = f in Ω , (5.63a)
∂xi ∂xj ∂xi
∂u
aij Ni = g on ∂Ω , (5.63b)
∂xj
and ∂Ω is of class C m+2 . Then u ∈ H m+2 (Ω) and satisfies the estimate
h i
kukH m+2 (Ω) ≤ C kf kH m (Ω) + kgkH m+0.5 (∂Ω) + kukL2 (Ω)
for some constant C depending only on m, Ω, and the coefficient of L.
5.10 Two-phase problems - Elliptic equations with

mixed type of boundary condition
In continuum mechanics, it is quite natural to consider multi-phase prob-

lems. For example, two immiscible fluids, such as water and oil, present
in the same fluid container. While the dynamics of fluids inside different
phases looks independent, they interact through the common interface thus
the behavior of the fluid could be quite complicated.
In the following, we propose a two-phase elliptic problem motivated by
the two-phase Hele-Shaw equation, and study the existence, uniqueness,
and regularity of the solution. Suppose that Ω1 and Ω2 are two disjoint
connected domain separated by a common boundary Γ, called the interface
between Ω1 and Ω2 , so that Ω = Ω1 ∪ Ω2 is connected (thus ∂Ω2 = ∂Ω ∪ Γ).
We call Ω1 and Ω2 the outer phase and the inner phase, respectively. We
consider
−div(b1 Dp1 ) = f1 in Ω1 , (5.64a)

−div(b2 Dp2 ) = f2 in Ω2 , (5.64b)
p1 = p2 on Γ, (5.64c)
(b1 Dp1 − b2 Dp2 ) · N = 0 on Γ, (5.64d)
p1 = 0 on ∂Ω , (5.64e)
where N on Γ is chosen to be a fixed orientation, and b1 , b2 are positive

functions with positive lower bounds, that is, b1 ≥ λ 0 and b2 ≥ λ 0 for some
constant λ 0 > 0.
Equation (5.64) might look over-constrained because it looks like two
separated elliptic equations with both Dirichlet and Neumann boundary
conditions. However, the following example indicates that (5.64) will not be
well-posed if imposing only (5.64c) or only (5.64d) as the interface condition.
§5.10 Two-phase problems 227
Example 5.58. Let Ω1 = (−2, −1) ∪ (1, 2) and Ω2 = (−1, 1). Then for all
a, b ∈ R, p1 and p2 given by
p1 (x) = x2 + ax + (2a − 4) 1(−2,−1) (x) + x2 + (a − 4)x + (4 − 2a) 1(1,2) (x) ,

p2 (x) = (a − 2)x + b ,
are solutions to
−p001 (x) = −2 for all x ∈ Ω1 ,

−p002 (x) = 0 for all x ∈ Ω2 ,
p10 (x) − p20 (x) = 0 at x = −1 and x = 1 ,
p1 (x) = 0 at x = −2 and x = 2 .
Similarly, for all a, c ∈ R , p1 and p2 given by
p1 (x) = x2 + ax + (2a − 4) 1(−2,−1) (x) + x2 + cx − (2c + 4) 1(1,2) (x) ,

a+c a−c−6
p2 (x) = − x+ ,
2 2
are solutions to
−p001 (x) = −2 for all x ∈ Ω1 ,

−p002 (x) = 0 for all x ∈ Ω2 ,
p1 (x) = p2 (x) at x = −1 and x = 1 ,
p1 (x) = 0 at x = −2 and x = 2 .
However, this is only one solution,
p1 (x) = x2 + 2x 1(−2,−1) (x) + x2 − 2x 1(1,2) (x) ,

p2 (x) = −1 ,
to the equation
−p001 (x) = −2 for all x ∈ Ω1 ,

−p002 (x) = 0 for all x ∈ Ω2 ,
p1 (x) = p2 (x) at x = −1 and x = 1 ,
p10 (x) + p20 (x) = 0 at x = −1 and x = 1 ,
p1 (x) = 0 at x = −2 and x = 2 .
5.10.1 The variational formulation and weak solutions
Suppose for now that there exists smooth solution p1 and p2 to the problem.
Testing (5.64a) against a test function ϕ ∈ H 1 (Ω2 ) and integrating by parts,
we find that
∂p1
Z Z Z
b1 Dp1 · Dϕ dx − b1 (τ1 ϕ) dS = f1 ϕ dx , (5.65)
Ω1 Γ ∂N1 Ω1
where N1 is the outward-pointing unit normal of Ω1 , and τ1 is the trace

operator on Ω1 . Similarly, testing (5.64b) against a test function ϕ ∈ H 1 (Ω1 )
so that ϕ = 0 on ∂Ω (ϕ might not vanish on Γ) and integrate by parts, we
find that
∂p2
Z Z Z
b2 Dp2 · Dϕ dx − b2 (τ2 ϕ) dS = f2 ϕ dx (5.66)
Ω2 Γ ∂N2 Ω1
where N2 is the outward-pointing unit normal of Ω2 , andτ2 is the trace

operator on Ω2 . Since N1 = −N2 on Γ,
∂p1 ∂p1
Z Z
b1 (τ1 ϕ) dS + b2 (τ1 ϕ) dS = 0 if τ1 ϕ = τ2 ϕ .
Γ ∂N1 Γ ∂N1
Therefore, for all ϕ ∈ H01 (Ω) , summing (5.65) and (5.65) together we obtain
that for all ϕ ∈ H01 (Ω),
Z Z Z Z
b1 Dp1 · Dϕ dx + b2 Dp2 · Dϕ dx = f1 ϕ dx + f2 ϕ dx .
Ω1 Ω2 Ω1 Ω2
Let p = p1 1Ω1 + p2 1Ω2 , b = b1 1Ω1 + b2 1Ω2 , f = f1 1Ω1 + f2 1Ω2 , then the

integral identity above implies that (5.64) can be rewritten as
−div(bDp) = f in Ω, (5.67a)
p=0 on ∂Ω . (5.67b)
The discussion above leads to the following
Definition 5.59. For any f ∈ H −1 (Ω), the variational form of (5.64) is

Z Z
b1 Dp1 · Dϕ dx + b2 Dp2 · Dϕ dx = hf, ϕiH01 (Ω) ∀ ϕ ∈ H01 (Ω) .
Ω1 Ω2
(p1 , p2 ) is called a weak solution to (5.64) if it satisfies the variational for-

mulation.
Since (5.64) can be rewritten as a one-phase Dirichlet problem (5.67),

by Theorem 5.38 there exists a unique weak solution p to the problem.
5.11. MAXIMUM PRINCIPLE 229
5.10.2 Regularity
In elliptic regularity, the coefficient aij are assumed to be C k (Ω) in order

to obtain u ∈ H k+1 (Ω). However, in this two-phase problem, b1 does not
have to be identical to b2 on the interface Γ. In other words, the coefficient
b in (5.67) might be discontinuous in Ω; thus the coefficient b will possess
limited regularity, say b ∈ L∞ (Ω). However, b1 and b2 can still be smooth
in Ω1 and Ω2 . This is the case that we are going to explore, and we have
the following
Theorem 5.60. Suppose that b1 ∈ C 1 (Ω1 ) and b2 ∈ C 1 (Ω2 ) have positive

lower bounds. Then for any f ∈ L2 (Ω) , the weak solution p ∈ H01 (Ω) to
(5.64) in fact belongs to the space
H 2 (Ω1 ) ∩ H 2 (Ω2 ) ≡ p ∈ H01 (Ω) p ∈ H 2 (Ω1 ) and p ∈ H 2 (Ω2 ) ,

and satisfies
kpkH 2 (Ω1 ) + kpkH 2 (Ω2 ) ≤ Ckf kL2 (Ω) .
5.11 Maximum principle
To study the maximum principle, we assume that the elliptic operator L is

of non-divergence form, that is,
∂2u ∂u
Lu = −aij + bi + cu .
∂xi ∂xj ∂xi
We remark here that L in divergence form can be rewritten as the non-

divergence form above if aij ∈ C 1 (Ω).
The following results generalize the maximum principle in Chapter 4.
Theorem 5.61 (Weak maximum principle). Assume u ∈ C 2 (Ω) ∩ C 0 (Ω)

and c ≥ 0 in Ω. Let u+ = max{u, 0} and u− = max{−u, 0}.
(i) If Lu ≤ 0 in Ω, then max u ≤ max u+ .

Ω ∂Ω
(ii) If Lu ≥ 0 in Ω, then min u ≤ − max u− .

Ω ∂Ω
Proof. Will be added later.

Lemma 5.62 (Hopf’s lemma). Assume u ∈ C 2 (Ω) ∩ C(Ω) and c ≡ 0 in Ω.

Suppose further that Lu ≤ 0 in Ω, and x0 ∈ ∂Ω is a point with the property
that
u(x0 ) > u(x) ∀x ∈ Ω,
and there exists an open ball B ⊆ Ω with x0 ∈ ∂B (this is call the interior
∂u
ball condition at x0 ). Then (x0 ) > 0, where N is the outward pointing
∂N
unit normal to Ω an x0 .
Proof. Assume that B = B(y, r) for some y ∈ Ω and r > 0. Define

2 2
v(x) = e−λ|x−y| − e−λr , x ∈ B(y, r)
for some λ > 0 to be chosen later. Then if A = [aij ] and b = (b1 , · · · , bn ),
∂2v ∂v
Lv = −aij + bi
∂xi ∂xj ∂xi
2
h i
= e−λ|x−y| aij − 4λ2 (xi − yi )(xj − yj ) + 2λδ ij − 2λbi (xi − yi )

2
≤ e−λ|x−y| − 4θλ2 |x − y|2 + 2λtr(A) + 2λ|b||x − y| .

By defining the open annular region R = B(y, r) − B(y, r/2), we have

2
Lv ≤ e−λ|x−y| − 4θλ2 |x − y|2 + 2λtr(A) + 2λ|b||x − y| ≤ 0

in R
if λ > 0 is large enough.

Now let w = u + v − u(x0 ). Then w ≤ 0 on ∂R, and Lw ≤ 0 in R; hence
the weak maximum principle (Theorem 5.61) implies that w ≤ 0 in R. On
the other hand, w(x0 ) = 0; thus
∂w ∂u ∂v
0≤ (x0 ) = (x0 ) + (x0 )
∂N ∂N ∂N
which implies
∂u ∂v |x0 − y|2 −λ|x0 −y|2 2
(x0 ) ≥ − (x0 ) = 2λ e = 2λre−λr > 0.
∂N ∂N r
Theorem 5.63 (Strong maximum principle). Suppose that Ω is a connected,
open, bounded domain of Rn , and u ∈ C 2 (Ω) ∩ C(Ω) and c ≡ 0 in Ω.
(i) If Lu ≤ 0 in Ω, and u attains its maximum over Ω at an interior point

x ∈ Ω, then u is constant within Ω.
5.12. EIGENVALUES AND EIGENFUNCTIONS 231
(ii) If Lu ≥ 0 in Ω, and u attains its minimum over Ω at an interior point

x ∈ Ω, then u is constant within Ω.
Proof. Let M = max u, A = {x ∈ Ω | u(x) = M }, and S = {x ∈ Ω | u(x) <

Ω
M }. If u 6≡ M , then S 6= ∅ and S is an open set . Choose a point y ∈ S such
that dist(y, A) < dist(y, ∂Ω), and let B denote the largest ball with center
y whose interior lies in S. Then for some x0 interior point of A, x0 ∈ ∂B.
∂u
Therefore, by the Hopf lemma, (x0 ) > 0, but this contradicts to the fact
∂N
that if Du(x0 ) = 0.
5.12 Eigenvalues and eigenfunctions
In this section, we consider the following eigenvalue problem
Lw = λw in Ω, (5.68a)
w=0 on ∂Ω , (5.68b)
where Ω is an open bounded smooth domain, and λ is called an eigenvalue

of L if there exists non-trivial w satisfying (5.68), in which case w is called
an eigenfunction of L.
For simplicity, we consider L with divergence form, and b = c = 0, that
is,
∂ ij ∂u
Lu = − a .
∂xi ∂xj
We assume that a ∈ C ∞ (Ω), aij = aji , and L is strictly elliptic in Ω.
Definition 5.64 (Spectrum). The collection of eigenvalues of L is called

the spectrum of L.
Theorem 5.65. Let Σ be the spectrum of L. Then Σ ⊆ R, and Σ is at most

countable. Let Σ = {λk }∞
k=1 , where the eigenvalue is counted according to
its multiplicity, then
0 < λ1 ≤ λ2 ≤ · · ·
and λk → ∞ as k → ∞. Finally, there exists an orthonormal basis {wk }∞

k=1
of L2 (Ω), where wk ∈ H01 (Ω) is an eigenfunction corresponding to λk , that
is,
Lwk = λk wk in Ω,
wk = 0 on ∂Ω .
Remark 5.66. By the elliptic regularity Theorem 5.56, wk ∈ C ∞ (Ω) for all
k ∈ N.
Proof. Let S = L−1 . Then S : L2 (Ω) → L2 (Ω) is a bounded, linear, compact

operator. If λ is an eigenvalue of L, then λ−1 is an eigenvalue of S. By
Theorem B.31, we know that σ(S), the spectrum of S, is at most countable,
and has 0 as the unique limit point.
We claim that S is symmetric; that is, for f, g ∈ L2 (Ω), we have that
(Sf, g)L2 (Ω) = (f, Sg)L2 (Ω) . With u = Sf and v = Sg,
Lu = f in Ω, Lv = g in Ω,
and
u=0 on ∂Ω , v=0 on ∂Ω ,
so that
∂v ∂u
Z
(Sf, g)L2 (Ω) = (u, Lv)L2 (Ω) = aij dx
Ω ∂xj ∂xi
∂u ∂v
Z
= aij dx = (Lu, v)L2 (Ω) = (f, Sg)L2 (Ω) .
Ω ∂xj ∂xi
The symmetry of S implies that Σ ⊆ R. The unique solvability of L for λ < 0
implies that Σ ⊆ R+ . Since σ(S) has 0 as the unique limit point, λk → ∞.
By Theorem B.38, the eigenfunctions of S make up an orthonormal basis
of L2 (Ω), and eigenfunctions of S are eigenfunctions of L as long as the
corresponding eigenvalues are not zero.
n o∞
Corollary 5.67. √wλk is an orthonormal basis of H01 (Ω), here the
k k=1
inner product of H01 (Ω) is defined by (u, v)H01 (Ω) = (Du, Dv)L2 (Ω) .
Proof. Since (wk , w` )H01 (Ω) = λk δk` , it suffices to show that if u ∈ H01 (Ω) so
that (u, wk )H01 (Ω) = 0 for all k ∈ N, then u = 0. However,
(u, wk )H01 (Ω) = λk (u, wk )L2 (Ω) ,
so (u, wk )H01 (Ω) = 0 implies that u is orthogonal to wk . The completeness of

{wk }∞
k=1 implies that u = 0.
§5.12 Eigenvalues and eigenfunctions 233
Definition 5.68. We call λ1 > 0 the principal eigenvalue of L.
Theorem 5.69. The principal eigenvalue λ1 is simple, and the correspond-

ing eigenfunction w1 has only one sign, that is, w1 > 0 in Ω or w1 < 0 in
Ω. Moreover,
nZ ∂u ∂u o
λ1 = min aij dx u ∈ H01 (Ω) , kukL2 (Ω) = 1 . (5.69)

Ω ∂xj ∂xi
Proof. We first prove (5.69). Since {wk }∞

k=1 is an orthonormal basis of
L2 (Ω), for a given u ∈ L2 (Ω) with kukL2 (Ω) = 1,
∞
X
u= dk wk , (5.70)
k=1
∞
d2k = kuk2L2 (Ω) = 1.
P
where dk = (u, wk )L2 (Ω) . By the Parseval identity,
k=1
ij ∂u ∂v
Z
Let B[u, v] ≡ a ij
dx. Since a is symmetric, B(·, ·) can be
Ω ∂xj ∂xi
viewed as an inner product of H01 (Ω), and { √wλk }∞k=1 forms an orthonormal
k
basis of H01 (Ω) with respect to this new inner product (because B(wk , wk ) =
λk kwk k2L2 (Ω) = λk and B(wk , w` ) = λk (wk , w` )L2 (Ω) = 0 if k 6= `). Moreover,
∞
X wk
u= µk √
k=1
λk
√
with µk = B(u, √wλk ). By (5.70) µk = dk λk ; hence
k
∞
X ∞
X
B[u, u] = µ2k = d2k λk ≥ λ1 .
k=1 k=1
Since B[w1 , w1 ] = λ1 , (5.69) is valid.

Claim: If u ∈ H01 (Ω) and kukL2 (Ω) = 1, then u is a weak solution of the
eigenvalue problem
Lu = λ1 u in Ω, (5.71a)
u=0 on ∂Ω , (5.71b)
if and only if B[u, u] = λ1 .

Proof of claim: The sufficiency is obvious. To prove the necessity, suppose
that u ∈ H01 (Ω), kukL2 (Ω) = 1 satisfies B[u, u] = λ1 . Let dk = (u, wk )L2 (Ω) ,
then
∞
X ∞
X ∞
X
λ1 d2k = λ1 = B[u, u] = d2k λk ⇒ d2k (λk − λ1 ) = 0 .
k=1 k=1 k=1
As a consequence, dk = 0 if λk > λ1 . Since λ1 has finite multiplicity,

m
X
u= (u, wk )wk
k=1
for some m ∈ N, where Lwk = λ1 wk for k = 1, · · · , m. But then Lu = λ1 u

which concludes the necessity. End of proof of claim.
Now let u ∈ H01 (Ω) with kukL2 (Ω) = 1 be a non-trivial weak solution of
(5.71). Then u+ and u− satisfy
( (
+ Du a.e. on {u ≥ 0} − 0 a.e. on {u ≥ 0}
Du = Du =
0 a.e. on {u ≤ 0} , −Du a.e. on {u ≤ 0} .
Moreover, B[u+ , u− ] = 0. Therefore,
λ1 = B[u, u] = B[u+ , u+ ] + B[u− , u− ]

by (5.69)
≥ λ1 ku+ k2L2 (Ω) + λ1 ku− k2L2 (Ω) = λ1 ;
hence the inequality above is in fact an equality. As a consequence,
B[u+ , u+ ] = λ1 ku+ k2L2 (Ω) and B[u− , u− ] = λ1 ku− k2L2 (Ω) ;
thus u+ and u− are both weak (in fact smooth) solutions to (5.71). In
particular, Lu+ ≥ 0 in Ω, and the strong maximum principle (Theorem
5.63) implies that u+ > 0 in Ω or u+ ≡ 0 in Ω. Same arguments apply to
u− , so u has only one sign in Ω.
Finally suppose
Z that u1 and u2 are two non-trivial weak solutions to
(5.71). Then uk dx 6= 0 for k = 1, 2. Therefore, there exists a constant
Ω
χ 6= 0 such that
Z
(u1 − χu2 )dx = 0 .
Ω
Since u1 −χu2 is again a weak solution to (5.71), it follows that u1 −χu2 ≡ 0

in Ω, hence the eigenvalue λ1 is simple.
5.13. EXERCISES 235
5.13 Exercises
Problem 5.1. The Lax-Milgram theorem cannot directly be applied to the

Robin problem (5.29) in the case that k is negative. This, however, does not
imply that there does not exist a solution to the Robin problem with k < 0.
This exercise is concerned with the solvability of the Robin problem in
du
the one-dimensional setting. For Ω = (a, b), we introduce the notation
dn
to mean
(
du a + b 0 u0 (x) if x = b ,
(x) = sgn x − u (x) =
dn 2 −u0 (x) if x = a .
Complete the following.
1. Find the unique solution to the Robin problem
−u00 (x) = −2 for all x ∈ (0, 1) ,

du
(x) − u(x) = 0 at x = 0 and x = 1 .
dn
2. Find the non-trivial solutions to the Robin problem
−u00 (x) = 0 for all x ∈ (0, 2) ,

du
(x) − u(x) = 0 at x = 0 and x = 2 .
dn
3. Show that there is no solution to the Robin problem
−u00 (x) = 6x for all x ∈ (0, 2) ,

du
(x) − u(x) = 0 at x = 0 and x = 2 .
dn

4. Next, suppose that Ω = x ∈ R2 |x| < 1 . Prove that for all f ∈

L2 (Ω) there exists a unique weak solution to
u − ∆u = f in Ω,
du
− ∆g u = 0 on ∂Ω .
dn
Problem 5.2. Let η = (η1 , η2 ) be a vector-valued function such that
(a) η1 : [0, 1] → [0, 1] is continuous, and η1 (0) = 0 and η1 (1) = 1;

(b) η1 : (0, 1) → (0, 1) is a diffeomorphism;
(c) η2 : [0, 1] → R is continuous;
(d) η2 : (0, 1) → R is differentiable.
Define g = η 0 · η 0 = |η 0 |2 , n = g −1/2 (−η20 , η10 ), and

0
H(η) = −g −1/2 g −1/2 η 0 · n .
Complete the following.
1. Show that H(η) = −g −1 (η 00 · n).
2. Let η be a family of functions satisfying (a)-(d) with η 0 (x) = (x, 0)

d
and η = u. Show that
d =0
d
L1 u ≡ H(η ) = −u002 .
d =0
This quantity is called the linearization of H(η) about η = η 0 .
3. Find a solution to L1 u = 1. Is this solution unique?
4. Let θ : (0, 1) → (0, 1) be a diffeomorphism. Show that
H(η ◦ θ) = H(η) ◦ θ .
Therefore, if η is a solution to H(η) = 1 , then η ◦ θ is also a solution

for every diffeomorphism θ. Consequently, the kernel of H is infinitely
dimensional.
5. Since η1 : (0, 1) → (0, 1) is a diffeomorphism, η1 has a unique inverse

η1−1 . Define h = η2 ◦ η1−1 . Then η ◦ η1−1 (x) = x, h(x) . Show that

−3/2 0
H(h) ≡ H(η ◦ η1−1 ) = 1 + |h0 |2 h00 = (1 + |h0 |2 )−1/2 h0 .
6. Let h0 be a given function on (0, 1) satisfying (c) and (d). Find the
linearization of H(h) about h = h0 . In other words, compute
d
L2 δh ≡ H(h )
d =0
d

given that h0 = h0 and h = δh .
d =0
§5.13 Exercise 237
7. Show that the solution to L2 δh = 1 is unique for any give h0 .
Problem 5.3. Let Ω ⊆ Rn denote an open, bounded, and smooth domain.

Let u : Ω → Rn denote a vector-valued function, and p : Ω → R be a scalar
function.
1. Find a unique solution to the PDE
u + Dp = 0 in Ω, (5.72a)
divu = 0 in Ω, (5.72b)
u·N=g on ∂Ω . (5.72c)
2. Find a unique solution to (5.72a,b) with (5.72c) replaced by the bound-

ary condition
p=g on ∂Ω .
3. Find an estimate for u and p in terms of the given data g for parts 1.
and 2.
4. Next, suppose that f ∈ H 1 (Ω) and that
u + Dp = f in Ω,
divu = 0 in Ω,
p = −∆g (u · N) on ∂Ω ,
where N denotes the outward unit normal to ∂Ω. Prove that there
exists a unique solution to this problem and find the estimates verified
by u and p.
Problem 5.4. Recall that
1
(Ω; R3 ) ≡ w ∈ H 1 (Ω; R3 ) divw = 0, w · N = 0 on ∂Ω .

Hdiv
We define the Leray projector P : H 1 (Ω; R3 ) → Hdiv

1 (Ω; R3 ) as follows: for
a vector u ∈ H 1 (Ω; R3 ), find w ∈ Hdiv

1 (Ω; R3 ) and p ∈ H 2 (Ω) such that
u = w + Dp. As long as the decomposition is unique, Pu ≡ w.
1. What is the (elliptic) equation that p must satisfy?

2. Is the decomposition u = w + Dp unique? In other words, is it possible

to find (w1 , p1 ) 6= (w2 , p2 ) such that w1 + Dp1 = w2 + Dp2 ?
3. Show that P : H 1 (Ω; R3 ) → Hdiv

1 (Ω; R3 ) is a continuous projection,
that is,
kPukH 1 3 ≤ CkukH 1 (Ω;R3 ) and P(Pu) = u ∀ u ∈ H 1 (Ω; R3 ) .

div (Ω;R )
Problem 5.5. In this problem, we examine if the specification of the diver-

gence and curl of a vector with periodic boundary conditions is sufficient to
uniquely determine the vector field.
Let u : T3 → R3 be a vector-valued function defined on T3 .
1. Show that curlcurlu = −∆u + Ddivu for all u ∈ C 2 (T3 ).
2. Show that
kDuk2L2 (T3 ) = kdivuk2L2 (T3 ) + kcurluk2L2 (T3 ) ∀ u ∈ H 1 (T3 ) , (5.74)
3
X
where kDuk2L2 (T3 ) ≡ kui ,j k2L2 (T3 ) .
i,j=1
3. Given f ∈ H −1 (T3 ) and g ∈ H −1 (T3 ; R3 ), find a vector field u ∈

L2 (T3 ; R3 ) with divu = f and curlu = g. Is this u unique? (Hint:
Write u in the form u = Dp + curlw for some p ∈ H 1 (T3 ) and w ∈
1 (T3 ; R3 ).)
Hdiv
Problem 5.6 (Calderon-Zygmund-type inequality). Let u ∈ H 3 (R3 ; R3 )

with divu = 0 and let ω = curlu. Then, with log+ |f | := log max(1, |f |),
h i
kDukL∞ (R3 ) ≤ C 1 + 1 + log+ kukH 3 (R3 ) kωkL∞ (R3 ) + kωkL2 (R3 ) .

This problem extends the basic elliptic estimate

h i
kukW 1,p (R3 ) ≤ Cp kcurlukLp (R3 ) + kdivukLp (R3 ) 1 < p < ∞,
which reads for the above u as
kukW 1,p (R3 ) ≤ Cp kωkLp (R3 ) 1 < p < ∞,

§5.13 Exercise 239
where the constant Cp → ∞ as p → ∞. The lemma shows that in the limit

as p → ∞, a logarithmic correction enters the estimate.
The proof of this result relies on the so-called Biot-Savart kernel which
is fundamental in fluid dynamics and electromagnetism. Since divu = 0, we
may introduce a vector potential ψ so that
u = −curlψ .
Then ω = curlu = ∆ψ and this implies that ψ = ∆−1 ω; thus, using the
fundamental solution for Poisson’s equation on R3 ,
1 ω(y)
Z
ψ(x) = − dy ,
4π R3 |x − y|
and hence
1 x−y
Z
u(x) = −curlψ(x) = − × ω(y)dy .
4π R3 |x − y|3
x
Definition 5.70. K(x)f = × f is called the Biot-Savart kernel, so
4π|x|3
that
Z Z
u(x) = K(x − y) ω(y)dy = K(y) ω(x − y)dy .
R3 R3
Use this definition to establish the result.
Hint. In computing Du(x), you must differentiate under the integral sign,
a process that leads to a singular integral that must be interpreted as a p.v.
integral, i.e.,
Z Z
K(y)Dω(x − y)dy := lim K(y)Dω(x − y)dy .
R3 δ→0 |y|≥δ
Problem 5.7. Show that for any p ∈ [1, n), there exist constants α ∈ (0, 1),
ν ∈ (0, 1], and a real symmetric n × n matrix a(x) = aij (x) satisfying the

uniform ellipticity condition
ν|ξ|2 ≤ aij (x)ξi ξj ≤ ν −1 |ξ|2 for all ξ ∈ Rn and x ∈ Rn ,
and such that the function u(x) ≡ 1 − |x|α satisfies
Lu ≡ aij u,ij = 0 a.e. in B(0, 1) ⊆ Rn ,
and its derivatives D2 u ∈ Lp (B(0, 1)). This problem shows that the maxi-
mum principle does not hold in general for uniformly elliptic equation with
discontinuous coefficients in the Sobolev class W 2,p , when p < n.
Problem 5.8. Let ψ : U ⊆ Rn → V ⊆ Rn be a diffeomorphism, and

f : V → R is a smooth function. Let x ∈ U and y = ψ(x) ∈ V be one
coordinate system on U and V. Show that
∂f ∂(f ◦ ψ)
◦ ψ = Aji ,
∂yi ∂xj
where A = (Dψ)−1 .
Hint: Recall the chain rule in the theory of partial derivatives of a function
of multiple variables.
Problem 5.9. Given f ∈ L2 (Ω). Let Kf ∈ H 2 (Ω) denote the unique

solution to the Neumann problem
(1 − ∆)u = f in Ω,
∂u
=0 on ∂Ω .
∂N
Show that K : L2 (Ω) → H 2 (Ω) is bounded, and K : L2 (Ω) → L2 (Ω) is
compact and symmetric. Construct an orthonormal basis {ek }∞ 2
k=1 in L (Ω)
which is orthogonal in H 1 (Ω) using the eigenfunctions of K.
Let u ∈ H 1 (Ω). Since u ∈ L2 (Ω), we may express u in terms of the basis
k
{e` }∞
P
`=1 . Suppose that uk = (u, e` )L2 (Ω) e` is the partial sum. Show that
`=1
uk → u in H 1 (Ω), and
kuk kH 1 (Ω) ≤ kukH 1 (Ω) . (5.75)
Suppose that u ∈ H 2 (Ω). In general uk does not converge to u since if so,

then
∂uk ∂u
0= → as k → ∞,
∂N ∂N
∂u
while does not have to vanish on ∂Ω. Now suppose that u ∈ H 2 (Ω)
∂N
∂u
satisfies = 0 on ∂Ω. Show that uk → u in H 2 (Ω), and there is a
∂N
constant C > 0 such that
kuk kH 2 (Ω) ≤ CkukH 2 (Ω) . (5.76)

Chapter 6
Linear Parabolic Equations
6.1 Spaces involving time
It is often important to identify functions u : Ω × [0, T ] → R with maps from

[0, T ] into a Banach space. For example, in the study of parabolic equations,
the space L2 (0, T ; H 1 (Ω)) plays an especially important role.
Suppose that X denotes an abstract Banach space with norm k · kX .
Definition 6.1. The space Lp (a, b; X) consists of functions u : (a, b) → X

so that u(t) ∈ X for almost all t ∈ (a, b). The Lp (a, b; X)-norm of u is given
as follows:
 hZ b i1
ku(·, t)kpX dt
p
if 1 ≤ p < ∞ ,



kukLp (a,b;X) ≡ a


 sup ku(·, t)kX if p = ∞ ,
t∈(a,b)
We say that Lp (a, b; X) is the space of Lp -integrable functions from (a, b)

into X.
Definition 6.2. The space C([a, b]; X) consists of functions u : [a, b] → X

continuously, so that u(t) ∈ X for all t ∈ (a, b). The C([a, b]; X)-norm of u
is given as follows:
kukC([a,b];X) = max ku(·, t)kX .

t∈[a,b]
For u ∈ C([a, b]; X), lim ku(·, t) − u(·, t0 )kX = 0 for all t0 ∈ [a, b]. In
t→t0
other words, C([a, b]; X) is the space of continuous functions from [a, b] into
X.
241
242 CHAPTER 6. Linear Parabolic Equations
Theorem 6.3. Lp (a, b; X) and C([a, b]; X) are Banach spaces for all p ∈
[1, ∞].
Proof. The proof is the same as the for the case that X = R.
Lemma 6.4. Let X be a given Banach space with dual X 0 . Suppose that
u, g ∈ L1 (a, b; X). Then, the following three conditions are equivalent
(1) u is a.e. equal to the anti-derivative of g; that is, for some ξ ∈ X

sometimes denoted by u(a) ,
Z t
u(t) = ξ + g(s) ds for almost all t ∈ [a, b] , (6.1)
a
(2) For each test function ϕ ∈ D(a, b),

Z b Z b dϕ
u(t)ϕ 0 (t) dt = − g(t)ϕ(t) dt ϕ0 =
a a dt
(3) For each η ∈ X 0 ,

d
hu, ηiX = hg, ηiX ,
dt
in the sense of distribution.
If (1)-(3) are satisfied, u is equal to a continuous function from [a, b] into

X.
Proof. It is easy to see that (1) implies (2) and (3), while the proof of (2)
implies (1) is similar to the proof of Theorem 2.112 and is left as an exercise.
It suffices to show that (3) implies (2).
Suppose that (3) holds and ϕ ∈ D(a, b), then by definition
Z b Z b
hu(t), ηiX ϕ 0 (t) dt = − hg(t), ηiX ϕ(t) dt
a a
or
DZ b Z b E
u(t)ϕ 0 (t) dt + g(t)ϕ(t) dt, η =0 ∀η ∈ X0
a a X
which implies (2).
Definition 6.5. The function g in (6.1) is called the weak time-derivative

of u, and is denoted by u 0 , ut or ∂t u.
6.1. SPACES INVOLVING TIME 243
The following corollary is the direct consequence of Lemma 6.4.
Corollary 6.6. If u 0 ∈ L1 (a, b; X), then u ∈ C([a, b]; X).
The notion of weak continuity plays a very important role in the analysis
of time-dependent PDE.
Definition 6.7. A function u ∈ L1 (a, b; X) is said to be weakly continuous

with values in X , denoted by u ∈ C([0, T ]; Xw ), if
hu, ϕiX ∈ C([0, T ]) ∀ϕ ∈ X0.
We use hu, ϕiX to denote the duality pairing between ϕ ∈ X 0 and u ∈ X.
Lemma 6.8. Let X and Y be two Banach spaces such that X ,→ Y (X

is continuously embedded in Y ) and X is dense in Y . If a function u ∈
L∞ (0, T ; X) and is weakly continuous with values in Y , then u is weakly
continuous with values in X.
Proof. Let ũ : R → X denote an extension of u to R satisfying ũ(t) =

u(t)χ[0,T ] (t), and define um = η 1 ∗ ũ where {η 1 }∞
m=1 denotes a sequence of
m m
standard mollifiers, and where the convolution operator is smoothing ũ in
time.
By definition, kum (t)kX ≤ kukL∞ (0,T ;X) for all t ∈ [0, T ], and since
u ∈ C([0, T ]; Yw ),
lim hum (t), ϕiX = hu(t), ϕiX ∀ t ∈ [0, T ] , ϕ ∈ Y 0 .

m→∞

Since hum (t), ϕiX ≤ kukL∞ (0,T ;X) kϕkX 0 for all m and t ∈ [0, T ], we find
that
∀ t ∈ [0, T ] , ϕ ∈ Y 0 .

hu(t), ϕiX = lim hum (t), ϕiX ≤ kukL∞ (0,T ;X) kϕkX 0
m→∞
This implies that the linear functional hu(t), ·iX is bounded on Y 0 . Since
Y 0 is dense in X 0 , this functional is also bounded on X 0 with

ku(t)kX = sup hu(t), ϕiX ≤ kukL∞ (0,T ;X) ∀ t ∈ [0, T ] .
kϕkX 0 =1
Now suppose that t0 ∈ [0, T ] and ϕ ∈ X 0 . Then for any given > 0, there
exists ψ ∈ Y 0 such that kϕ − ψkX 0 < and we find that

hu(t) − u(t0 ), ϕiX ≤ hu(t) − u(t0 ), ψiX + hu(t) − u(t0 ), ϕ − ψiX

≤ hu(t) − u(t0 ), ψiY + 2kukL∞ (0,T ;X) .
Therefore, by the weak continuity of u with values in Y ,
lim sup |hu(t) − u(t0 ), ϕiX | ≤ 2kukL∞ (0,T ;X)

t→t0
which guarantees the weak continuity of u with values in X.
Lemma 6.9. Suppose u ∈ L2 (0, T ; H), with u 0 ∈ L2 (0, T ; H 0 ) for some

Hilbert space H so that H ,→L2 (Ω) ⊆ H 0 (for example, H can be H 1 (Ω) or
H01 (Ω)). Then
d
(1) kuk2L2 (Ω) = 2hu0 , uiH in D0 ([0, T ]), where u0 is the weak time-derivative
dt
of u; and
(2) u ∈ C([0, T ]; L2 (Ω)), and

1 h i
max ku(t)kL2 (Ω) ≤ 1 + kukL2 (0,T ;H) + ku 0 kL2 (0,T ;H 0 ) . (6.2)
t∈[0,T ] T
Proof. Let ϕ ∈ C0∞ ([0, T ]), and uk ∈ C ∞ (Ω × [0, T ]) so that uk → u in

L2 (0, T ; H) and uk0 → u 0 in L2 (0, T ; H 0 ) as k → ∞. Then (1) follows from
the following equality:
Dd E Z TZ
0
kuk2L2 (Ω) , ϕ 2
|u(x, t)|2 ϕ 0 (t)dxdt

= − kukL2 (Ω) , ϕ = −
dt 0 Ω
Z TZ
= − lim |uk (x, t)|2 ϕ 0 (t)dxdt
k→∞ 0 Ω
Z TZ Z T
0
= lim 2uk (x, t)uk (x, t)ϕ(t)dxdt = 2hu 0 , uiH ϕdt .
k→∞ 0 Ω 0
To prove (2), we first show that u : [0, T ] → L2 (Ω) is continuous; that

is, we show that for each t0 ∈ [0, T ], lim ku(t) − u(t0 )k2L2 (Ω) = 0. However,
t→t0
lim ku(t)k2L2 (Ω) = ku(t0 )k2L2 (Ω) by (1) , and u(t) − u(t0 ), ϕ L2 (Ω) → 0 as

t→t0
t → t0 for all ϕ ∈ L2 (Ω) by Corollary 6.6 and Lemma 6.8 (with X = L2 (Ω)
6.1. SPACES INVOLVING TIME 245
and Y = H 0 ); hence
lim ku(t) − u(t0 )k2L2 (Ω)

t→t0
h i
= lim ku(t)k2L2 (Ω) + ku(t0 )k2L2 (Ω) − 2 u(t), u(t0 ) L2 (Ω) = 0 .

t→t0
Finally, by (1) we find that for almost all t, s ∈ [0, T ],

Z t
2 2
ku(t)kL2 (Ω) − ku(s)kL2 (Ω) = 2hu(s), u 0 (s)iH ds
s
≤ kuk2L2 (0,T ;H) + ku 0 k2L2 (0,T ;H 0 ) .
Integrate in s from 0 to T , the equality above implies

1 T
Z
ku(t)k2L2 (Ω) ≤ ku(s)k2L2 (Ω) ds + kuk2L2 (0,T ;H) + ku 0 k2L2 (0,T ;H 0 )
T 0

1 h i
≤ 1+ kuk2L2 (0,T ;H) + ku 0 k2L2 (0,T ;H 0 ) .
T
Remark 6.10. The application of Corollary 6.6 and Lemma 6.8 for ob-
taining the continuity of u can be re-sentenced as “convergence in a weaker
sense (continuity in the weak sense) and the convergence of norms imply
convergence in a stronger sense (continuity in norm)”. This is a general-
ized version of the following well-known fact: if fn → f a.e. or weakly in
Lp (Ω), and kfn kLp (Ω) → kf kLp (Ω) , then fn → f in Lp (Ω).
Corollary 6.11. Let Ω ⊆ Rn be an open, bounded domain with smooth

boundary ∂Ω. Suppose that u ∈ L2 (0, T ; H m+1 (Ω)) and u 0 ∈ L2 (0, T ; H m−1 (Ω))
for some non-negative integer m. Then u ∈ C([0, T ]; H m (Ω)) and
h i
max ku(t)kH m (Ω) ≤ C kukL2 (0,T ;H m+1 (Ω)) + ku 0 kL2 (0,T ;H m−1 (Ω)) . (6.3)
t∈[0,T ]
6.1.1 Compactness theorems
By the Rellich theorem, H 1 (Ω)⊂⊂L2 (Ω) if Ω is bounded and smooth. How-

ever, it is not the case that L2 (0, T ; H 1 (Ω))⊂⊂L2 (0, T ; L2 (Ω)) since a se-
quence of functions uk (t) ∈ L2 (0, T ) clearly also belong to L2 (0, T ; H 1 (Ω))
but might not have convergent subsequence in L2 (0, T ; L2 (Ω)). In this sub-
section, we are concerned with pre-compact subsets of L2 (0, T ; H k (Ω)) or
L2 (0, T ; H01 (Ω)). .
Lemma 6.12. Let X0 , X, X1 be Banach spaces such that
X0 ⊂⊂X ,→X1 . (6.4)
Then for each > 0, there exists a constant C (and the spaces) such that
kxkX ≤ kxkX0 + C kxkX1 ∀ x ∈ X0 .
Proof. Suppose the contrary. Then for some fixed > 0, there exists xk ∈ X0
such that
kxk kX ≥ kxk kX0 + kkxk kX1 .
Let yk = xk /kxk kX . Then the inequality above implies kkyk kX1 +kyk kX0 ≤
1. In particular, yk is bounded in X0 , and yk converges to 0 in X1 . By
X0 ⊂⊂X, there exists a subsequence ykj that converges to y in X. Since
yk converges to 0 in X1 , y must be 0 (this is the place where we use the
continuous embedding of X into X1 ), while the strong convergence of yk in
X implies that kykX = 1 which is a contradiction.
Theorem 6.13. Let X0 , X, X1 be Banach spaces satisfying (6.4), and

X0 , X1 are reflexive. Let T > 0 be a fixed finite number, p0 , p1 ∈ (1, ∞),
and Y be the function space consisting of all u ∈ Lp0 (0, T ; X0 ) such that
u 0 ∈ Lp1 (0, T ; X1 ); that is,
n o
Y ≡ u ∈ L (0, T ; X0 ) u 0 ∈ Lp1 (0, T ; X1 )
p0
with norm kykY = kykLp0 (0,T ;X0 ) + kykLp1 (0,T ;X1 ) (which makes Y a Banach
space). Then
Y ⊂⊂Lp0 (0, T ; X) .
Proof. Let uk be a bounded sequence in Y . Since p0 , p1 ∈ (1, ∞) and X0 ,

X1 are reflexive, the spaces L2 (0, T ; X0 ) and L2 (0, T ; X1 ) are also reflexive;
hence Y is reflexive as well. By Banach-Alaoglu theorem, there exists ukj
that converges weakly to u ∈ Y ; that is,
ukj * u in Lp0 (0, T ; X0 ) and ukj t * ut in Lp1 (0, T ; X1 ) .

6.2. SECOND-ORDER PARABOLIC EQUATIONS 247
We show that this weak limit u is in fact the strong limit of a subsequence
of ukj in Lp0 (0, T ; X).
Let vj = ukj − u. The weak convergence of vjt in Lp1 (0, T ; X1 ) implies
that the function vj is a bounded sequence in C([0, T ]; X1 ) since by Lemma
6.4
Z t2 Z t2
kvj (t2 ) − vj (t1 )kX1 = vjt (s) ds ≤ kvjt (s)kX1 ds

t1 X1 t1
p1 −1
≤ |t2 − t1 | p1 kvjt kLp1 (0,T ;X1 ) .
The proof of the Arzela-Ascoli Theorem then also shows that there is a
further subsequence vj` (t) converges to 0 in C([0, T ]; X1 ). Hence vj` → 0 in
Lp0 (0, T ; X1 ).
Finally, by Lemma 6.12,
kvj` kLp0 (0,T ;X) ≤ kvj` kLp0 (0,T ;X0 ) + C kvj` kLp0 (0,T ;X1 ) . (6.5)
Therefore, the uniform boundedness of vj` in Lp0 (0, T ; X0 ) implies
lim kvj` kLp0 (0,T ;X) ≤ C ;

`→∞
hence vj` → 0 in Lp0 (0, T ; X0 ) or ukj` → u in Lp0 (0, T ; X0 ).
6.2 Second-order parabolic equations
We consider the following initial-boundary value problem
ut + Lu = f in Ω × (0, T ) , (6.6a)
u = u0 on Ω × {t = 0} , (6.6b)
with either the Dirichlet boundary condition
u=0 on ∂Ω × (0, T ) (6.6c)
or the Neumann boundary condition
∂u
aij Ni = g on ∂Ω × (0, T ) , (6.6d)
∂xj
where L is a time-dependent elliptic operator defined by
∂ ij ∂u ∂u
(Lu)(x, t) = − a (x, t) (x, t) + bi (x, t) (x, t) (6.7)
∂xi ∂xj ∂xi
+ c(x, t)u(x, t)
and the matrix A = [aij ]. We note that since now problem is time dependent,
the coefficients a, b and c may depend on t as well.
∂
Definition 6.14. We say that the partial differential operator + L is
∂t
uniformly parabolic if there exists a constant λ 0 > 0 such that aij (x, t)ξj ξi ≥
λ 0 |ξ|2 for all (x, t) ∈ Ω × (0, T ), ξ ∈ Rn .
6.2.1 Weak solutions
Assume that aij , bi , c ∈ L∞ (Ω × (0, T )), f ∈ L2 (0, T ; L2 (Ω)), and u0 ∈

L2 (Ω).
Definition 6.15 (Weak solutions with Dirichlet boundary conditions). A

function u ∈ L2 (0, T ; H01 (Ω)) with ut ∈ L2 (0, T ; H −1 (Ω)) is said to be a weak
solution of the parabolic initial-boundary value problem (6.6a,b,c) provided
that

ut (t), ϕ H 1 (Ω) + B u(t), ϕ = f (t), ϕ L2 (Ω)
0
(6.8)
∀ ϕ ∈ H01 (Ω), a.e. t ∈ (0, T ),
and
u(0) = u0 , (6.9)

where B u(t), ϕ is defined by
∂u ∂ϕ
Z h
aij (x, t)

B u(t), ϕ ≡ (x, t) (x)
Ω ∂xj ∂xi
(6.10)
i ∂u i
+ b (x, t) (x, t)ϕ(x) + c(x, t)u(x, t)ϕ(x) dx .
∂xi
Equation (6.8) is called the variational formulation of the parabolic problem

(6.6a,b,c).
Definition 6.16 (Weak solutions with Neumann boundary conditions). A

function u ∈ L2 (0, T ; H 1 (Ω)) with ut ∈ L2 (0, T ; H 1 (Ω)0 ) is said to be a weak
§6.2 Second-order parabolic equations 249
solution of the parabolic initial-boundary value problem (6.6a,b,d) provided

that u satsifies both (6.9) and

ut (t), ϕ H 1 (Ω) + B u(t), ϕ = f (t), ϕ L2 (Ω) + g(t), ϕ H 0.5 (∂Ω)
0
(6.11)
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ),

where B u(t), ϕ is defined by (6.10). Equation (6.11) is called the varia-
tional formulation of the parabolic problem (6.6a,b,d).
Remark 6.17. By Lemma 6.9, a weak solution u ∈ C([0, T ]; L2 (Ω)), so

condition (6.9) makes sense; in other words, the temporal trace at t = 0 is
well-defined.
We proceed to give the detailed construction of weak solutions for the

case of the Neumann boundary conditions (6.6d). The argument for the
case of the Dirichlet boundary condition (6.6c) is similar, and perhaps even
easier.
6.2.2 Existence of weak solutions - Galerkin schemes
Let {ek }∞ 2 1
k=1 be an orthonormal basis of L (Ω) which is orthogonal in H (Ω)
(given by Problem 5.9 in Chapter 5, for example). We define
k
X
uk (x, t) = dk` (t)e` (t)
`=1
such that the coefficients dk (t) are solutions of the following system:

ukt (t), ϕ L2 (Ω) + B uk (t), ϕ = f (t), ϕ L2 (Ω) + g(t), ϕ H 0.5 (∂Ω)
(6.12)
∀ ϕ ∈ span(e1 , · · · , ek ) .
The function uk (x, t) should be thought of as a finite-dimensional ap-

proximation of u(x, t); in fact, it is the L2 (Ω) projection onto the first k
orthonormal basis vectors {e1 , ... , el }.
The variational problem (6.12) is a coupled system of linear ordinary
differential equations for the vector dk (t) given by
dkm0 (t) + M`m (t)dk` (t) = (f (t), em )L2 (Ω) + g(t), em H 0.5 (∂Ω) ≡ Fm (t) ,

dk` (0) = (u0 , e` )L2 (Ω) ,

where the time-dependent matrix M`m (t) is defined by

∂e` ∂em
Z h
M`m (t) = aij (x, t) (x) (x)
Ω ∂xj ∂xi
∂e` i
+ bi (x, t) (x)em (x) + c(x, t)e` (x)em (x) dx .
∂xi
This system of linear ordinary differential equations can be written as dk 0 (t)+
M̃(t)dk (t) = F (t); M̃(t) = MT (t) is a k × k matrix, and dk (t), F (t) are vec-
tors.
Since a, b, c ∈ L∞ ((0, T ); L∞ (Ω)), the existence of a unique Lipschitz
continuous functions dk is guaranteed by the fundamental theorem of ODE,
since the matrix M̃ ∈ L∞ (0, T ). To be more precise, by the contraction
mapping theorem, for each k ∈ N, there exists a 0 < Tk ≤ T such that the
solution dk (t) satisfies
Z t Z t
k k k
d (t) = d (0) − M̃(s)d (s) ds + F (s) ds ∀ t ∈ [0, Tk ] .
0 0
However, since the ODE is linear, Tk = T .

To obtain a weak solution to equation (6.6a,b,d), it is natural to consider
the limit of uk as k → ∞. To guarantee the existence of a weak limit, it is
necessary for us to obtain uniform estimates for the sequences uk and ukt ,
with bounds that are independent of k.
Uniform estimates for uk
Since (6.12) holds for all ϕ in the span of e1 , · · · , ek , we may set
ϕ = uk
in (6.12) since it is linear combination of {e` }k`=1 . By doing so, we find that
1d ∂uk ∂uk
Z Z
2 ij
kuk (t)kL2 (Ω) + a (x, t) (x, t) (x, t)dx= f (x, t)uk (x, t)dx
2 dt Ω ∂xj ∂xi Ω
∂u
Z h i
k
bi (x, t) (x, t)uk (x, t) + c(x, t)u2k (x, t) dx .

+ g(t), uk (t) H 0.5 (∂Ω) −
Ω ∂xi
By the uniform parabolicity, the equality above suggests that
1d
Z
2 2
kuk (t)kL2 (Ω) + λ 0 kDuk (t)kL2 (Ω) ≤ f (x, t)uk (x, t)dx
2 dt Ω
∂uk
Z h i
bi (x, t) (x, t)uk (x, t) + c(x, t)u2k (x, t) dx .

+ g(t), uk (t) H 0.5 (∂Ω) −
Ω ∂xi
Hölder’s and Young’s inequalities further imply that

1d
kuk (t)k2L2 (Ω) + λ 0 kDuk (t)k2L2 (Ω) ≤ kuk (t)kL2 (Ω) ×
2 dt h i
× kf (t)kL2 (Ω) + kb(t)kL∞ (Ω) kDuk (t)kL2 (Ω) + kc(t)kL∞ (Ω) kuk (t)kL2 (Ω)
λ0 h 1 1 λ0 i
≤ kDuk (t)k2L2 (Ω) + kb(t)k2L∞ (Ω) + kc(t)kL∞ (Ω) + + kuk (t)k2L2 (Ω)
2 λ0 2 4
1 C
+ kf (t)k2L2 (Ω) + kg(t)k2H −0.5 (∂Ω) ;
2 λ0
hence
d 2C
kuk (t)k2L2 (Ω) + λ 0 kDuk (t)k2L2 (Ω) ≤ kf (t)k2L2 (Ω) + kg(t)k2H −0.5 (∂Ω)
dt λ0
h 2 λ0 i
+ kb(t)k2L∞ (Ω) + 2kc(t)kL∞ (Ω) + 1 + kuk (t)k2L2 (Ω) .
λ0 2
| {z }
≡K(t)
By the generalized Gronwall inequality (see Problem 3.8 in Chapter 3),

h 2C i
kuk (t)k2L2 (Ω) ≤ kuk (0)k2L2 (Ω) + kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H −0.5 (∂Ω)) ×
λ0
Z t
× exp K(s) ds
0
h i
≤ C ku0 k2L2 (Ω) + kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H −0.5 (∂Ω)) ,

for some constant C = C kbkL∞ (Ω×(0,T )) , kckL∞ (Ω×(0,T )) , λ 0 , T , where the
last inequality follows from the Bessel inequality stating that kuk (0)kL2 (Ω) ≤
ku0 kL2 (Ω) . As a consequence, dk exists on [0, T ], and
Z t
2
kuk (t)kL2 (Ω) + λ 0 kDuk (s)k2L2 (Ω) ds
i (6.13)
0
h
2 2 2
≤ C ku0 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H −0.5 (∂Ω)) .
In other words, uk exists and is uniformly bounded in L2 (0, T ; H 1 (Ω)).
The estimates for ukt
To estimate kukt kH 1 (Ω)0 , we need to find the supremum of hukt , ϕiH 1 (Ω)
over functions ϕ ∈ H 1 (Ω) with kϕkH 1 (Ω) = 1. For each ϕ ∈ H 1 (Ω) with
kϕkH 1 (Ω) = 1, we can decompose ϕ so that ϕ = ϕ1 +ϕ2 , where ϕ1 belongs to
the span of e1 , · · · , ek and (ϕ1 , ϕ2 )H 1 (Ω) = 0. Then kϕ1 kH 1 (Ω) ≤ kϕkH 1 (Ω) =
1.
Using the relation

ukt (t), ϕ1 L2 (Ω)
+ B uk (t), ϕ1 = f (t), ϕ1 L2 (Ω) + g(t), ϕ1 H 0.5 (∂Ω)
together with the fact that ukt belongs to the span of e1 , · · · , ek , we find
that
hukt (t), ϕiH 1 (Ω) = (ukt (t), ϕ)L2 (Ω) = (ukt (t), ϕ1 )L2 (Ω)

= (f (t), ϕ1 )L2 (Ω) + g(t), ϕ1 H 0.5 (∂Ω) − B uk (t), ϕ1 .

The expression for B uk (t), ϕ then shows that

B uk (t), ϕ1
h i
≤ ka(t)kL∞ (Ω) + kb(t)kL∞ (Ω) + kc(t)kL∞ (Ω) kuk (t)kH 1 (Ω) kϕ1 kH 1 (Ω) .
As a consequence,

kukt kH 1 (Ω)0 = sup hukt (t), ϕiH 1 (Ω) ≤ kf (t)kL2 (Ω) + kg(t)kH −0.5 (∂Ω)
kϕkH 1 (Ω) =1
h i
+ ka(t)kL∞ (Ω) + kb(t)kL∞ (Ω) + kc(t)kL∞ (Ω) kuk (t)kH 1 (Ω) ;
hence
Z T h i
kukt (t)k2H 1 (Ω)0 dt ≤ 4 kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H −0.5 (∂Ω))
0
Z Th i
+8 ka(t)k2L∞ (Ω) + kb(t)k2L∞ (Ω) + kc(t)k2L∞ (Ω) kuk (t)k2H 1 (Ω) dt .
0
By (6.13), the inequality above implies that

Z T h i
kukt (t)k2H 1 (Ω)0 dt ≤ C |u0 k2L2 (Ω) + kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H −0.5 (∂Ω))
0
for some constant C = C kakL∞ (Ω×(0,T )) , kbkL∞ (Ω×(0,T )) , kckL∞ (Ω×(0,T )) , T, λ 0 );

thus ukt is uniformly bounded in L2 (0, T ; H 1 (Ω)0 ).
Since L2 (0, T ; H 1 (Ω)) and L2 (0, T ; H 1 (Ω)0 ) are both Hilbert spaces, by
the Banach-Alaoglu Theorem A.10 there exists a subsequence kj of k such
that ukj and ukj t converge weakly in L2 (0, T ; H 1 (Ω)) and L2 (0, T ; H 1 (Ω)0 ),
respectively. Let u be the weak limit of ukj ; then the weak limit of ukj t must
be ut .
N
d˜` (t)e` (x) for some N ≤ kj , d˜` ∈ C 1 ([0, T ]). Since
P
Let v(x, t) =
`=1
v(t) ∈ span(e1 , · · · , ekj ), (6.12) implies that

ukj t (t), v(t) H 1 (Ω) + B ukj (t), ϕ = f (t), v(t) L2 (Ω) + g(t), v(t) H 0.5 (∂Ω) .
Integrating the equality above in time over [0, T ], the definition of weak
convergence in L2 (0, T ; H 1 (Ω)) and L2 (0, T ; H 1 (Ω)0 ) shows that
Z Th

i
ut (t), v(t) H 1 (Ω) + B u(t), v(t) dt
0
Z Th i

= f (t), v(t) L2 (Ω) + g(t), v(t) H 0.5 (∂Ω) dt .
0
N
d˜` (t)e` (x), is dense in L2 (0, T ; H 1 (Ω)), the equality
P
Since v, of the form
`=1
above implies that
Z Th Z Th

i
ut (t), v(t) H 1 (Ω) + B u(t), v(t) dt = f (t), v(t) L2 (Ω)
0
i
0 (6.14)
+ g(t), v(t) H 0.5 (∂Ω) dt ∀ v ∈ L (0, T ; H 1 (Ω)) .
2

Finally, letting v(x, t) = χ[a,b] (t)ϕ(x) for some ϕ ∈ H 1 (Ω) in (6.14), we

find that for 0 ≤ a < b ≤ T ,
Z bh

i
ut (t), ϕ H 1 (Ω) + B u(t), ϕ dt
a
Z bh i
f (t), ϕ L2 (Ω) + g(t), ϕ H 0.5 (∂Ω) dt ∀ ϕ ∈ H 1 (Ω),

=
a
(6.15)
and the variational formulation (6.11) follows from the Lebesgue differenti-
ation theorem.
Remark 6.18. From the discussion above, we know that (6.11) and (6.14)
are equivalent, so (6.14) can be thought as an alternative variational formu-
lation for the parabolic initial-boundary value problem (6.6a,b,d). Similarly,
an alternative variational formulation for equation (6.6a,b,c) is given by
Z Th

i
ut (t), v(t) H 1 (Ω) + B u(t), v(t) dt
0
Z T
∀ v ∈ L2 (0, T ; H01 (Ω)) .

= f (t), v(t) L2 (Ω) dt
0
(6.16)
In order to prove that u is indeed a weak solution, it remains to prove

that u(0) = u0 . We note that while by construction, uk (0) → u0 in L2 (Ω),
it is not necessary that u(0) = u0 .
Let ζ ∈ C 1 [0, T ] be such that ζ(0) = 1 and ζ(T ) = 0. Multiplying

(6.12) by ζ(t) and integrating over the time interval (0, T ), we find that for
all ϕ ∈ span(e1 , · · · , ekj ),
Z T h i
ukj t (t), ζ(t)ϕ L2 (Ω)
+ B ukj (t), ζ(t)ϕ dt
0
Z T h
i
= f (t), ζ(t)ϕ L2 (Ω) + g(t), ζ(t)ϕ H 0.5 (∂Ω) dt ,
0
(6.17)
and the use of v(x, t) = ζ(t)ϕ(x) (which belongs to L2 (0, T ; H 1 (Ω))) in (6.14)
implies that
Z Th

i
ut (t), ζ(t)ϕ L2 (Ω) + B u(t), ζ(t)ϕ dt
0
Z T h
i
= f (t), ζ(t)ϕ L2 (Ω) + g(t), ζ(t)ϕ H 0.5 (∂Ω) dt ,
0
(6.18)
for all ϕ ∈ H 1 (Ω). Integrating-by-parts in time on the first term of the
left-hand side of (6.17), we find that
Z Th i
B ukj (t), ζ(t)ϕ − ukj (t), ζ 0 (t)ϕ L2 (Ω) dt (6.19)

ukj (0), ϕ H 1 (Ω) +
0
Z Th i

= f (t), ϕ L2 (Ω) + g(t), ζ(t)ϕ H 0.5 (∂Ω) dt ∀ ϕ ∈ span(e1 , · · · , ekj ) ,
0
and similarly from (6.18),

Z Th i
B u(t), ζ(t)ϕ − u(t), ζ 0 (t)ϕ L2 (Ω) dt

u(0), ϕ L2 (Ω) + (6.20)
0
Z Th i
f (t), ϕ L2 (Ω) + g(t), ζ(t)ϕ H 0.5 (∂Ω) dt ∀ ϕ ∈ H 1 (Ω).

=
0
Passing (6.19) to the limit, since ukj (0) → u0 in L2 (Ω), we obtain that
T Z
B u(t), ζ(t) − hu(t), ζ 0 (t)ϕiH 1 (Ω) dt

(u0 , ϕ)L2 (Ω) + (6.21)
0
Z Th i
f (t), ϕ L2 (Ω) + g(t), ζ(t)ϕ H 0.5 (∂Ω) dt ∀ ϕ ∈ H 1 (Ω).

=
0

The comparison of (6.20) and (6.21) suggests that (u0 , ϕ)L2 (Ω) = u(0), ϕ L2 (Ω)
for all ϕ ∈ H 1 (Ω); hence u(0) = u0 a.e. This proves the existence of a weak
solution to equation (6.6a,b,d).
6.2.3 Uniqueness of the weak solution
Assume that u1 and u2 are two weak solutions. Let w = u1 − u2 . Then by

(6.14) w satisfies
Z T h
i
∀ v ∈ L2 (0, T ; H 1 (Ω)) ,

wt (t), v(t) H 1 (Ω) + B w(t), v(t) dt = 0
0
w=0 at t = 0 .
Since w ∈ L2 (0, T ; H 1 (Ω)), we may use it in the equality above, and similar
to the estimates that we obtained for uk , we find that
t=T Z T
kw(t)k2L2 (Ω) + λ0 kDw(t)k2L2 (Ω) dt

t=0 0
h kbk iZ T
L∞ (0,T ;L∞ (Ω))
≤ + 2kckL∞ (0,T ;L2 (Ω)) kw(t)k2L2 (Ω) dt .
2λ 0 0
The Gronwall inequality then implies that kw(t)k2L2 (Ω) = 0 since w(0) = 0,
hence the uniqueness of the weak solution to equation (6.6a,b,d).
We have established the following
Theorem 6.19. There exists a unique weak solution u to the parabolic

initial-boundary value problem (6.6a,b,c) (resp. (6.6a,b,d)) satisfying
max ku(t)kL2 (Ω) + kukL2 (0,T ;H 1 (Ω)) + kut kL2 (0,T ;H −1 (Ω))
t∈[0,T ]
h i
≤ C ku0 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω))
 
max ku(t)kL2 (Ω) + kukL2 (0,T ;H 1 (Ω)) + kut kL2 (0,T ;H 1 (Ω)0 )
t∈[0,T ]
resp.
 
h i
≤ C ku0 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H −0.5 (∂Ω))
(6.22)
for some constant C depending on λ 0 , T , kakL∞ (0,T ;L∞ (Ω)) , kbkL∞ (0,T ;L∞ (Ω))
and kckL∞ (0,T ;L∞ (Ω)) .
Remark 6.20. The condition f ∈ L2 (0, T ; L2 (Ω)) can be further weakened

to f ∈ L2 (0, T ; H 1 (Ω)0 ), and the right-hand side of the weak formulation
(6.11) is then replaced by hf (t), ϕiH01 (Ω) or hf (t), ϕiH 1 (Ω) . The estimates are

almost identical, and the right-hand side of (6.22) will then be C ku0 kL2 (Ω) +

kf kL2 (0,T ;H −1 (Ω)) or C ku0 kL2 (Ω) + kf kL2 (0,T ;H 1 (Ω)0 ) .
Remark 6.21. Unlike the elliptic case, no matter how large b and c are,
the solution to the parabolic initial-boundary value problem (6.6a,b,c) or
(6.6a,b,d) always exists and is unique. For the problem with the Neumann
boundary conditions, the initial condition is used to rule out the non-zero
constant functions in the kernel of the differential operator ∂t + L.
6.2.4 Regularity
The regularity result for the parabolic initial-boundary value problem relies
on elliptic regularity, and estimates for the time-derivatives of the solution.
For example, suppose that we could prove that ut (t) ∈ L2 (Ω); then according
to the elliptic regularity theorem for the Neumann boundary condition, we
conclude that u(t) ∈ H 2 (Ω).
Improving the regularity of ut requires more regularity of the initial
data and boundary data. To proceed, we assume that u0 ∈ H 1 (Ω), g ∈
L2 (0, T ; H 0.5 (∂Ω)) with gt ∈ L2 (0, T ; H −0.5 (∂Ω)), and the matrix A = aij

is symmetric; that is, aij = aji . Setting ϕ = ukt in (6.12), we find that
∂uk ∂ukt
Z
kukt (t)k2L2 (Ω) + aij (x, t)
(x, t) (x, t)dx
Ω ∂xj ∂xi
Z Z
= f (x, t)ukt (x, t)dx + g(x, t)ukt (x, t)dS
Ω ∂Ω
∂uk
Z h i
− bi (x, t) (x, t)ukt (x, t) + c(x, t)uk (x, t)ukt (x, t) dx .
Ω ∂xi
Since we assume that a is symmetric,
∂uk ∂ukt 1d ∂uk ∂uk

Z Z
aij (x, t) (x, t) (x, t)dx = aij (x, t) (x, t) (x, t)dx
Ω ∂xj ∂xi 2 dt Ω ∂xj ∂xi
1 ∂uk ∂uk
Z
− aij
t (x, t) (x, t) (x, t) dx . (6.23)
2 Ω ∂xj ∂xi
3 2 1 2
Hölder’s inequality together with Young’s inequality, ab ≤ a + b , then
2 6
leads to
1d ∂uk ∂uk
Z
kukt (t)k2L2 (Ω) + aij (x, t)
(x, t) (x, t)dx
2 dt
Ω ∂xj ∂xi
1 3 h
≤ kat (t)kL∞ (Ω) kDuk (t)k2L2 (Ω) + kb(t)k2L∞ (Ω) kDuk (t)k2L2 (Ω)
2 2 i 1
+ kc(t)k2L∞ (Ω) kuk (t)k2L2 (Ω) + kf (t)k2L2 (Ω) + kukt (t)k2L2 (Ω)
Z 2
+ g(x, t)ukt (x, t)dS .
∂Ω
(6.24)
For the last term on the right-hand side, we note that integrating by parts
in time implies that
Z tZ
g(x, s)ukt (x, s) dSds
0 ∂Ω
Z s=t Z t

= g(x, s)uk (x, s)dS − gt (s), uk (s) H 0.5 (∂Ω) dt

∂Ω s=0 0
≤ Ckg(t)kH −0.5 (∂Ω) kuk (t)kH 1 (Ω) + Ckg(0)kH −0.5 (∂Ω) kuk (0)kH 1 (Ω)
Z t
+C kgt (s)kH −0.5 (∂Ω) kuk (s)kH 1 (Ω) ds
0
h 1
≤ C ku0 k2H 1 (Ω) + kg(0)k2H −0.5 (∂Ω) + kg(t)k2H −0.5 (∂Ω) (6.25)
4λ 0
i λ
0
+ kgt k2L2 (0,T ;H −0.5 (∂Ω)) + kf kL2 (0,T ;L2 (Ω)) + kuk (t)k2H 1 (Ω) .
2
Moreover, by the interpolation inequality

h i
max kg(t)kL2 (∂Ω) ≤ C kgkL2 (0,T ;H 0.5 (∂Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω)) , (6.26)
t∈[0,T ]
(6.25) implies that

Z tZ h
g(x, s)ukt (x, s) dSds ≤ C ku0 k2H 1 (Ω) + kf k2L2 (0,T ;L2 (Ω))
0 ∂Ω
i λ
0
+ kgk2L2 (0,T ;H 0.5 (∂Ω)) + kgt k2L2 (0,T ;H −0.5 (∂Ω)) + kuk (t)k2H 1 (Ω) .
2
(6.27)
As a consequence, integrating (6.24) over the time interval (0, t), by the
parabolicity and (6.13) we find that
t
λ0
Z
kukt (s)k2L2 (Ω) ds + kDuk (t)k2L2 (Ω)
0 2
h i
≤ C ku0 k2H 1 (Ω) + kf k2L2 (Ω) + kgk2L2 (0,T ;H 0.5 (∂Ω)) + kgt k2L2 (0,T ;H −0.5 (∂Ω)
for some constant C = C kakL∞ (Ω×(0,T )) , kbkL∞ (Ω×(0,T )) , kckL∞ (Ω×(0,T )) , T, λ 0 ).

Therefore, the weak limit ut belongs to L2 (0, T ; L2 (Ω)) and satisfies
h
kut k2L2 (0,T ;L2 (Ω)) ≤ C ku0 k2H 1 (Ω) + kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H 0.5 (∂Ω))
i
+ kgt k2L2 (0,T ;H −0.5 (∂Ω)) .
Next, we rewrite the weak formulation (6.11) as
B(u, ϕ) = (f − ut , ϕ)L2 (Ω) + (g, ϕ)L2 (∂Ω) ∀ ϕ ∈ H 1 (Ω), t ∈ (0, T ) a.e. .
For almost every instant of time, we thus have a weak solution u to the ellip-
∂u
tic problem Lu = f − ut with the Neumann boundary condition aij Ni =
∂xj
g. By elliptic regularity,
h i
ku(t)kH 2 (Ω) ≤ C(t) kf (t) − ut (t)kL2 (Ω) + kgkH 0.5 (∂Ω) + ku(t)kL2 (Ω)

for some constant C(t) = C ka(t)kC 1 (Ω) , kb(t)kC 1 (Ω) , kc(t)kC 1 (Ω) , λ 0 . Thus
for a, b, c ∈ C 1 (Ω × [0, T ]),
h
kukL2 (0,T ;H 2 (Ω)) ≤ C ku0 kH 1 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H 0.5 (∂Ω))
i
+ kgt kL2 (0,T ;H −0.5 (∂Ω)) .
We conclude with the following
Theorem 6.22. Assume that ∂t +L is uniformly parabolic, ∂Ω is of class C 2 ,

aij , bi , c ∈ C 1 (Ω × [0, T ]), and aij is symmetric. Then for any u0 ∈ H 1 (Ω),

f ∈ L2 (0, T ; L2 (Ω)), g ∈ L2 (0, T ; H 0.5 (∂Ω)) with gt ∈ L2 (0, T ; H −0.5 (∂Ω)),

the unique weak solution u to (6.6a,b,d) in fact belongs to L2 (0, T ; H 2 (Ω)) ∩
C([0, T ]; H 1 (Ω)) with ut ∈ L2 (0, T ; L2 (Ω)) and satisfies
max ku(t)kL2 (Ω) + kut kL2 (0,T ;L2 (Ω)) + kukL2 (0,T ;H 2 (Ω))
t∈[0,T ]
h
≤ C ku0 kH 1 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H 0.5 (∂Ω)) (6.28)
i
+ kgt kL2 (0,T ;H −0.5 (∂Ω)) .
Remark 6.23. For the Dirichlet problem (6.6a,b,c), the same argument
leading to Theorem 6.22 can be applied to obtain ut ∈ L2 (0, T ; L2 (Ω)) with
the additional assumption u0 = 0 on ∂Ω. The reason for requiring that u0

vanishes on ∂Ω is for the validity of the following inequality
kDuk (0)kL2 (Ω) ≤ Cku0 kH 1 (Ω) (6.29)
for some constant C independent of k, and u0 = 0 on ∂Ω is called the

first-order compatibility condition for the parabolic initial-boundary problem
(6.6a,b,c). We note that if u0 does not vanish on ∂Ω, (6.29) is no longer
true (see Problem 6.6 for the counterexample). Thus similar to Theorem
6.22 we have the following
Theorem 6.24. Assume that ∂t +L is uniformly parabolic, ∂Ω is of class C 2 ,

aij , bi , c ∈ C 1 (Ω × [0, T ]), and aij is symmetric. Then for any u0 ∈ H01 (Ω),

f ∈ L2 (0, T ; L2 (Ω)), the unique weak solution u to (6.6a,b,c) in fact belongs

to L2 (0, T ; H 2 (Ω)) ∩ C([0, T ]; H01 (Ω)) with ut ∈ L2 (0, T ; L2 (Ω)) and satisfies
max ku(t)kL2 (Ω) + kut kL2 (0,T ;L2 (Ω)) + kukL2 (0,T ;H 2 (Ω))
t∈[0,T ]
(6.30)
≤ C ku0 kH 1 (Ω) + kf kL2 (0,T ;L2 (Ω)) .
6.2.5 Higher regularity and compatibility conditions
To address the issues of obtaining higher regularity, we first focus the dis-
cussion on one of the simplest parabolic equations
ut − ∆u = f in Ω × (0, T ) , (6.31a)
∂u
=g on ∂Ω × (0, T ) , (6.31b)
∂N
u = u0 on Ω × {t = 0} . (6.31c)
Suppose in addition that f ∈ L2 (0, T ; H 1 (Ω)) with ft ∈ L2 (0, T ; H 1 (Ω)0 ),

g ∈ L2 (0, T ; H 1.5 (∂Ω)), and u0 ∈ H 2 (Ω). A solution of the time-differentiated
problem
wt − ∆w = ft in Ω × (0, T ) , (6.32a)
∂w
= gt on ∂Ω × (0, T ) , (6.32b)
∂N
w = f (0) + ∆u0 ≡ u1 on Ω × {t = 0} , (6.32c)
can be related to the solution of (6.31) via the fundamental theorem of

calculus; that is,
Z t
u(x, t) = u0 (x) + w(x, s) ds ∀x ∈ Ω. (6.33)
0
(We have assumed sufficient regularity on u0 such that w|t=0 ∈ L2 (Ω).)

To validate (6.33), we consider the weak form of (6.32):

wt (t), ϕ H 1 (Ω) + Dw(t), Dϕ L2 (Ω) = ft (t), ϕ H 1 (Ω) + gt (t), ϕ H 0.5 (∂Ω)
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ).
Integrating the equation above in time, we find that for all ϕ ∈ H 1 (Ω),

Z t
w(t), ϕ H 1 (Ω) + Dw(s) ds, Dϕ 2
L (Ω)

0

= w(0), ϕ H 1 (Ω) + f (t) − f (0), ϕ H 1 (Ω) + g(t) − g(0), ϕ H 0.5 (∂Ω)

= f (t), ϕ L2 (Ω) + (∆u0 , ϕ)L2 (Ω) + g(t) − g(0), ϕ L2 (∂Ω)

= f (t), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω) − (Du0 , Dϕ)L2 (Ω)
∂u0
Z h i
+ − g(0) ϕ dS ;
∂Ω ∂N
hence for all ϕ ∈ H 1 (Ω),

Z t Z t

u0 + w(s)ds t , ϕ + D u0 + w(s) ds , Dϕ
0 L2 (Ω) 0 L2 (Ω)
∂u0
Z h i

= f (t), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω) + − g(0) ϕ dS .
∂Ω ∂N
Z t
∂u0
This shows that u0 + w(s) ds is a weak solution to (6.31) if = g(0),
0 ∂N
and by the uniqueness of the weak solution, we indeed have (6.33). The
condition
∂u0
= g(0) on ∂Ω (6.34)
∂N
is called the first-order compatibility condition for (6.31).
Now since ft ∈ L2 (0, T ; H 1 (Ω)0 ) and u0 ∈ H 2 (Ω), the weak solution
w = ut to (6.32) belongs to L2 (0, T ; H 1 (Ω)) with wt ∈ L2 (0, T ; H(Ω)0 )
satisfying
max kut (t)kL2 (Ω) + kut kL2 (0,T ;H 1 (Ω)) + kutt kL2 (0,T ;H(Ω)0 )
t∈[0,T ]
h i
≤ C ku0 kH 2 (Ω) + kf (0)kL2 (Ω) + kft kL2 (0,T ;H(Ω)0 ) + kgt kL2 (0,T ;H −0.5 (∂Ω)) .
Moreover, −∆u = f − ut in the weak sense, so elliptic regularity implies

that
h i
ku(t)kH 3 (Ω) ≤ C kf (t)kH 1 (Ω) + kut (t)kH 1 (Ω) + kgkH 1.5 (∂Ω)
and hence
h i
max kut (t)kL2 (Ω) + ku(t)kH 2 (Ω) + kukL2 (0,T ;H 3 (Ω)) + kut kL2 (0,T ;H 1 (Ω))
t∈[0,T ]
h
+ kutt kL2 (0,T ;H(Ω)0 ) ≤ C ku0 kH 2 (Ω) + kf kL2 (0,T ;H 1 (Ω)) + kft kL2 (0,T ;H(Ω)0 )
i
+ kgkL2 (0,T ;H 1.5 (∂Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ). Here we use (6.2) to estimate kf (0)kL2 (Ω) .
On the other hand, if u ∈ L2 (0, T ; H 3 (Ω)) with ut ∈ L2 (0, T ; H 1 (Ω)),
∂u
then u ∈ C([0, T ]; H 2 (Ω)) thus Du ∈ C([0, T ]; H 1 (Ω)) which implies ∈
∂N
∂u
C([0, T ]; H 0.5 (∂Ω)). Since by construction = g on ∂Ω, we have
∂N
∂u0 ∂u(t)
= lim = g(0) on ∂Ω .
∂N t→0 ∂N
Therefore, (6.34) is also a necessary condition for (6.31) to have a solution
u ∈ L2 (0, T ; H 3 (Ω)) with ut ∈ L2 (0, T ; H 1 (Ω)).
Suppose further that u0 ∈ H 3 (Ω) and that f ∈ L2 (0, T ; H 2 (Ω)), ft ∈
L2 (0, T ; L2 (Ω)), g ∈ L2 (0, T ; H 2.5 (∂Ω)), gt ∈ L2 (0, T ; H 0.5 (∂Ω)) and gtt ∈
L2 (0, T ; H −0.5 (∂Ω)). Then the solution w to the time-differentiated problem
(6.32) belongs to L2 (0, T ; H 2 (Ω)) ∩ L∞ (0, T ; H 1 (Ω)) by Theorem 6.22. No
additional compatibility condition is required to guarantee the additional
gain of regularity of w. However, in order to make a connection between w
and u via (6.33), the first-order compatibility condition is still needed.
Suppose even further that u0 ∈ H 4 (Ω) and that f ∈ L2 (0, T ; H 3 (Ω)),
ft ∈ L2 (0, T ; H 1 (Ω)), and ftt ∈ L2 (0, T ; H 1 (Ω)0 ), as well as g ∈ L2 (0, T ; H 3.5 (∂Ω)),
gt ∈ L2 (0, T ; H 1.5 (∂Ω)) and gtt ∈ L2 (0, T ; H −0.5 (∂Ω)). In this case, w, ft ,
gt and u1 play the role of u, f , g and u0 in (6.31); hence to understand the
regularity of the solution, it is necessary to time-differentiate (6.32) once
more. We find that the solution u to (6.31) belongs to L2 (0, T ; H 5 (Ω)) as
long as one additional condition on the initial data is satisfied:

∂w(0) ∂ f (0) + ∆u0
= = gt (0) on ∂Ω . (6.35)
∂N ∂N
The condition (6.35) is the first-order compatibility condition for the time-
differentiated problem (6.32), and is called the second-order compatibility
condition for (6.31).
In the following, we summarize what we have concluded in a table.
k The requirement for u∈L2 (0, T ;H k (Ω)) in addition to u0 ∈H k−1 (Ω)
1 f ∈L2 (0, T ;H 1 (Ω)0 ), g ∈L2 (0, T ;H −0.5(∂Ω)).
2 f ∈L2 (0, T ; L2 (Ω)), g ∈L2 (0, T ;H 0.5(∂Ω)), gt ∈L2 (0, T ;H −0.5(∂Ω)).
f ∈L2 (0, T ;H 1 (Ω)), ft ∈L2 (0, T ;H 1 (Ω)0 ), g ∈L2 (0, T ;H 1.5(∂Ω)),

3
gt ∈L2 (0, T ;H −0.5(∂Ω)), 1st-order compatibility condition.
f ∈L2 (0, T ;H 2 (Ω)), ft ∈L2 (0, T ; L2 (Ω)), g ∈L2 (0, T ;H 2.5(∂Ω)),

4 gt ∈L2 (0, T ;H 0.5(∂Ω)), gtt ∈L2 (0, T ;H −0.5(∂Ω)),
1st-order compatibility condition.
f ∈L2 (0, T ;H 3 (Ω)), ft ∈L2 (0, T ;H 1 (Ω)), ftt ∈L2 (0, T ;H 1 (Ω)0 ),
5 g ∈L2 (0, T ;H 3.5(∂Ω)), gt ∈L2 (0, T ;H 1.5(∂Ω)), gtt ∈L2 (0, T ;H −0.5(∂Ω)),
1st-order and 2nd-order compatibility conditions.
The discuss above motivates the following
Definition 6.25 (Compatible time-differentiated problems). The k-th com-

patible time-differentiated version of the parabolic initial-boundary value prob-
lem (6.6a,b,c) (resp. (6.6a,b,d)) consists of the following:
∂ku
1. a parabolic equation for wk ≡ obtained from time-differentiating
∂tk
(6.6a) k times ;
2. an initial condition defined by
∂ k−1
uk ≡ wk (0) = − (f − Lu) ; (6.36)
∂tk−1 t=0
3. a Dirichlet (resp. Neumann) boundary condition defined by

k−1
∂kg
ij ∂wk
X k ∂ k−` aij ∂w`

wk = 0 resp. a Ni = k − Ni on ∂Ω.
∂xj ∂t ` ∂tk−` ∂xj
`=0
(6.37)
We note that when considering equation (6.6a,b,d), the corresponding

boundary condition for the k-th compatible time-differentiated problem is
∂u
obtained from time-differentiating the boundary condition aij Ni = g k
∂xj
times.
Definition 6.26 (Compatibility conditions). For k ∈ N ∪ {0}, the (k + 1)th-

order compatibility condition for the parabolic initial-boundary value problem
(6.6a,b,c) (resp. (6.6a,b,d)) is given by τ uk = 0 on ∂Ω
k−1
∂uk ∂kg X k ∂ k−` aij ∂u`

resp. aij (0) Ni = k (0) − Ni (0) on ∂Ω ,
∂xj ∂t ` ∂tk−` ∂xj
`=0
where u` , ` = 1, · · · , k, is the initial condition of the `th compatible time-

differentiated problem for equation (6.6a,b,c) (resp. (6.6a,b,d)) defined in
(6.36).
Finally, for the purpose of defining the regularity of the solution and the
k (T ; Ω) and V k (T ; Ω)
forcing functions, we introduce the function spaces VD N
that consist of space-time dependent functions so that one time derivative
scales like two space derivatives. To be more concrete, for k ≥ 1,
VDk (T ; Ω)
n h k + 1 io
≡ u ∈ L (0, T ; H (Ω)) ∂tj u ∈ L2 (0, T ; H k−2j (Ω)) if 0 ≤ j ≤
2 k
,

2
[ k+1
2
]
k∂tj ukL2 (0,T ;H k−2j (Ω)) , and
X
equipped with norm kukV k (T ;Ω) =
D
j=0
VNk (T ; Ω)
n hki
≡ u ∈ L2 (0, T ; H k (Ω)) ∂tj u ∈ L2 (0, T ; H k−2j (Ω)) if 0 ≤ j ≤ ,

2
h k i [ k ]+1 o
k−2 ∂t 2 u ∈ L2 (0, T ; H 1 (Ω)0 ) ,
2
equipped with norm

k
[2]
[ k ]+1
h k i
k∂tj ukL2 (0,T ;H k−2j (Ω)) + k − 2
X
kukV k (T ;Ω) = k∂t 2 ukL2 (0,T ;H 1 (Ω)0 ) .
N 2
j=0
k (T ; Ω) and V k (T ; Ω) are the same when k is even.

Note that VD N
k+0.5
For the Neumann problem, we also need the space V∂Ω (T ) for defining
the regularity of the boundary data g: for k ∈ N ∪ {0, −1},
k+0.5
V∂Ω (T )
n hk + 1i
≡ g ∈L2 (0, T ; H k+0.5 (∂Ω)) ∂tj g ∈L2 (0, T ; H k+0.5−2j (∂Ω)) if 0 ≤ j ≤

o 2
k/2+1 2 −0.5
∂t g ∈ L (0, T ; H (∂Ω)) .
equipped with norm
[ k+1
2
]
k∂tj gkL2 (0,T ;H k+0.5−2j (∂Ω))
X
kgkV k+0.5 (T ) =
∂Ω
j=0
[ k+3
h k + 1 i ]
+ k+1−2 k∂t 2
gkL2 (0,T ;H −0.5 (∂Ω)) .
2
Having defined the function spaces, we can state the main
Theorem 6.27. Assume that ∂Ω is of class C m+2 , ∂t + L is uniformly

parabolic, aij , bi , c ∈ C m+1 (Ω × [0, T ]), and aij is symmetric. Then for

any u0 ∈ H m+1 (Ω) ∩ H01 (Ω) (resp. H m+1 (Ω)), f ∈ VDm (T ; Ω) (resp. f ∈
m+0.5
VNm (T ; Ω), g ∈ V∂Ω (T )), the unique weak solution u to the parabolic
initial-boundary value problem (6.6a,b,c) (resp. (6.6a,b,d)) in fact belongs to
VDm+2 (T ; Ω) (resp. VNm+2 (T ; Ω)) and satisfies
h i
kukV m+2 (T ;Ω) ≤C ku0 kH m+1 (Ω) + kf kVDm (T ;Ω)
D
h i (6.38)
resp. kukV m+2 (T ;Ω) ≤C ku0 kH m+1 (Ω) + kf kVNm (T ;Ω) + kgkV m+0.5 (T ) ,
N ∂Ω
hmi
provided that the compatibility conditions are valid up to + 1 -th (resp.
hm + 1i 2
-th) order.
2
6.3 Parabolic equations with special forcing func-

tions
We will explore a kind of parabolic initial-boundary value problem which

involves special kind of forcing functions and is useful in many real world
applications (and one of the applications will be seen in Chapter 7). The
§6.3 Parabolic equations with special forcing functions 265
equation we are going to study is
∂ ij ∂u ∂ ij ∂h
ut − a =f+ a in Ω × (0, T ) , (6.39a)
∂u ∂h
aij Ni = g − aij Ni on ∂Ω × (0, T ) , (6.39b)
∂xj ∂xj
u = u0 in Ω × {t = 0} , (6.39c)
where aij , ∈ L∞ (Ω × (0, T )), u0 ∈ L2 (Ω), f ∈ L2 (0, T ; L2 (Ω)) and g ∈

−0.5
V∂Ω (T ) as before, and in addition we assume that aij ∈ L∞ (Ω×(0, T )) and
h ∈ L2 (0, T ; H 1 (Ω)). Here we set bi = c = 0 in L to simplify the discussion.
At the first sight, (6.39) does not look very different from the parabolic
problem (6.6a,b,d) if we treat (aij h,j ),i as the forcing function and (6.39b)
is merely a non-homogeneous Neumann boundary condition. This is exactly
true when h possesses better regularity, say for example h ∈ L2 (0, T ; H 2 (Ω));
∂h
however, with h ∈ L2 (0, T ; H 1 (Ω)), the boundary data g−aij Ni does not
∂xj
make sense, so requires a little special treatment. Moreover, this problem is
of special interest especially when study the regularity theory for hyperbolic
equations.
Testing (6.39a) against a test function ϕ ∈ H 1 (Ω) results in the following
Definition 6.28. A function u ∈ L2 (0, T ; H 1 (Ω)) with ut ∈ L2 (0, T ; H 1 (Ω)0 )

is said to be a weak solution to (6.39) if
Z

ut (t), ϕ H 1 (Ω) + B u(t), ϕ = f (t), ϕ L2 (Ω) − B h(t), ϕ + g(t)ϕ dS
∂Ω
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ) , (6.40)
and u(0) = u0 , where B(u, ϕ) is defined by (6.10) with bi = c = 0, and

B(h, ϕ) is defined similarly with aij replacing aij . (6.40) is called the vari-
ational formulation of the parabolic equation (6.39).
For f, g, h with given regularity, (6.40) makes perfect sense, and the
existence of the unique weak solution to (6.39) follows exactly the same
fashion as before. In fact, since

B(h, ϕ) ≤ CkDhkL2 (Ω) kDϕkL2 (Ω) ,
by Young’s inequality, the finite dimensional approximation uk satisfies

1d λ0
kuk k2L2 (Ω) + kDuk k2L2 (Ω)
2 dt 2
1 Ch i 1 λ0
≤ kf k2L2 (Ω) + kDhk2L2 (Ω) + kgk2H −0.5 (∂Ω) + + kuk k2L2 (Ω)
2 λ0 2 4
which by Gronwall’s inequality gives uniform bounds for uk in L2 (0, T ; H 1 (Ω))
and ukt in L2 (0, T ; H 1 (Ω)0 ); thus the weak limit provides the unique weak
solution to (6.39).
Now suppose in addition that u0 ∈ H 1 (Ω), aij , aij ∈ C 1 (Ω × [0, T ]),
[aij ] is symmetric, g ∈ L2 (0, T ; H 0.5 (∂Ω)) and h ∈ L2 (0, T ; H 2 (Ω)). Then
Theorem 6.22 implies that u ∈ VN 2 (T ; Ω) and satisfies
Z Th i
2
max ku(t)kH 1 (Ω) + ku(t)k2H 2 (Ω) + kut (t)k2L2 (Ω) dt
t∈[0,T ] 0
h i
≤ C ku0 kH 1 (Ω) + kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H 0.5 (∂Ω)) + khk2L2 (0,T ;H 2 (Ω)) .
2
Suppose further that u0 ∈ H 2 (Ω), aij , aij ∈ C 2 (Ω × [0, T ]), f ∈ V 1 (T ; Ω),

1.5 (T ), and h ∈ L2 (0, T ; H 3 (Ω)) with h ∈ L2 (0, T ; H 1 (Ω)). For better
g ∈ V∂Ω t
regularity of u, we need to study the time differentiated problem

∂ ij ∂w ∂ ij ∂ht ∂h ∂u
wt − a = ft + a + aij
t + aij
t in Ω×(0, T ),
∂xi ∂xj ∂xi ∂xj ∂xj ∂xj
∂w
h ∂ht ∂h ∂u i
aij Ni = gt − aij + aij
t + aij
t Ni on ∂Ω×(0, T ),
∂xj ∂xj ∂xj ∂xj
∂ ij ∂u0 ∂ ij ∂h(0)
w = f (0) − a − a on Ω×{t = 0}.
By Corollary 6.11, h ∈ C([0, T ]; H 2 (Ω)) thus h(0) ∈ H 2 (Ω) which implies

that w(0) ∈ L2 (Ω). Therefore, the time-differentiated problem is in the form
of (6.39), and possesses a unique weak solution satisfying
Z Th i
2
max kw(t)kL2 (Ω)+ kw(t)k2H 1 (Ω)+kwt (t)k2H 1 (Ω)0 dt
t∈[0,T ] 0
h
≤C kf (0)k2L2 (Ω)+kh(0)k2H 2 (Ω)+ku0 k2H 2 (Ω)+kft k2L2 (0,T ;H 1 (Ω)0 )+kuk2L2 (0,T ;H 1 (Ω))
i
+khk2L2 (0,T ;H 1 (Ω))+kht k2L2 (0,T ;H 1 (Ω))+kgt k2L2 (0,T ;H −0.5 (∂Ω))
h
≤C ku0 k2H 2 (Ω)+kf k2L2 (0,T ;H 1 (Ω))+kft k2L2 (0,T ;H 1 (Ω)0 )+khk2L2 (0,T ;H 3 (Ω))
(6.41)
i
+kht k2L2 (0,T ;H 1 (Ω))+kgt k2L2 (0,T ;H −0.5 (∂Ω)) .
6.4. EXERCISES 267
As long as the first-order compatibility condition

∂u0 ∂h(0)
aij (0) Ni = g(0) − aij (0) Ni on ∂Ω × (0, T )
∂xj ∂xj
holds, (6.33) is satisfied and (6.41) implies that

h i
kuk2V 3 (T ;Ω) ≤ C ku0 k2H 2 (Ω) + kf k2V 1 (T ;Ω) + kgk2V 1.5 (T ) + khk2V 3 (T ;Ω) .
N N ∂Ω N
Inductively, by Theorem 6.27 we establish the following
Theorem 6.29. Assume that ∂Ω is of class C m+2 , ∂t + L is uniformly

parabolic, aij , aij , bi , c ∈ C m+1 (Ω × [0, T ]), and aij is symmetric. Then for

m (T ; Ω), g ∈ V m+0.5 (T ), h ∈ V m+2 (T ; Ω), the

any u0 ∈ H m+1 (Ω), f ∈ VN ∂Ω N
unique weak solution u to the parabolic initial-boundary value problem
∂ ij ∂h
ut + Lu = f + a in Ω × (0, T ) ,
∂xi ∂xj
∂u ∂h
aij Ni = g − aij Ni on ∂Ω × (0, T ) ,
∂xj ∂xj
u = u0 in Ω × {t = 0} ,
m+2 m+2
in fact belongs to VN (T ; Ω) (resp. VN (T ; Ω)) and satisfies
h
kukV m+2 (T ;Ω) ≤ C ku0 kH m+1 (Ω) + kf kVNm (T ;Ω) + kgkV m+0.5 (T )
N ∂Ω
i (6.42)
+ khkV m+2 (T ;Ω) ,
N
hm + 1i
provided that the compatibility conditions are valid up to -th order.
2
6.4 Exercises
Problem 6.1. Assume that uj * u in L2 (0, T ; H01 (Ω)) and that ∂t uk * v

in L2 (0, T ; H −1 (Ω)). Prove that v = ∂t u. (Hint. Let φ ∈ C01 (0, T ), and
Z T Z T
1
w ∈ H0 (Ω). Then h∂t uk , φwidt = − huk , φwidt.)
0 0
Problem 6.2. Let H be a Hilbert space, and u, ut ∈ L2 (0, T ; H). Show that
u ∈ C([0, T ]; H) and
1 h i
max ku(t)k2H ≤ 1 + kuk2L2 (0,T ;H) + kut k2L2 (0,T ;H) . (6.43)
t∈[0,T ] T
Problem 6.3. Let g ∈ C([0, T ]; L2 (Ω)) and u be such that ut ∈ L2 (0, T ; H 2 (Ω))
and
−κ∆ut − ∆u = g on Ω × [0, T ].
Then, independently of κ > 0,
k∆ukL∞ (0,T ;L2 (Ω)) ≤ kgkL∞ (0,T ;L2 (Ω)) + k∆u(0)kL2 (Ω) .
Problem 6.4. Let Ω = (0, 1)2 and let
1
(Ω) := u ∈ H 1 (Ω) u = 0 on {y = 0} ∩ {x = 0} ∩ {x = 1}

HB
Prove the existence of a weak solution u ∈ L2 (0, T ; HB

1 (Ω) ∩ H 1 (ω)) to the
parabolic problem
ut − ∆u = f in Ω × (0, T ] ,
u=0 on {x = 0} ∩ {x = 1} ∩ {y = 0} × (0, T ] ,
uy = uxx on (ω := {y = 1}) × (0, T ] ,
u = u0 on Ω × {t = 0} ,
where f ∈ L2 (0, T ; L2 (Ω) and u0 ∈ L2 (Ω).

Discuss the requirements on the data for regularity of solutions.
Problem 6.5. Let Ω ⊂ Rd for d > 1, and consider the time-dependent

Stokes problem for the velocity vector u = (u1 , ..., ud ) and the scalar pressure
function p, given by
ut − ∆u + Dp = f in Ω,
divu = 0 in Ω, (6.44)
u=0 on ∂Ω
where f ∈ L2 (0, T ; L2 (Ω; Rd )). Let
1
(Ω) = u ∈ H01 (Ω; Rd ) divu = 0 .

H0,div
Prove that there exists a unique weak solution u ∈ L2 (0, T ; H0,div

1 (Ω)) to the
Stokes problem.
Discuss the requirements on the data for regularity of solutions.
§6.4 Exercise 269
r
2
Problem 6.6. Let e` = sin `x, and ẽ` = `−1 e` . From Chapter 3 we
π
know that {e` }∞
`=1 is an orthonormal basis of L2 (0, π). Let the inner product
on H01 (0, π) be defined by (u, v)H01 (0,π) = (u0 , v 0 )L2 (0,π) . By Corollary 5.67
{ẽ` }∞ 1
`=1 is an orthonormal basis of H0 (0, π). Show that
(u, e` )L2 (0,π) = `−1 (u 0 , ẽ`0 )L2 (0,π) ∀ u ∈ H 1 (0, π);
k k
(u, ẽ` )H01 (0,π) ẽ` converges to u in H 1 (0, π)
P P
thus uk = (u, e` )L2 (0,π) e` =
`=1 `=1
for all u ∈ H01 (0, π).
Let u = 1 be a constant function. Then
k
4 X 4 sin(2` + 1)x
u2k+1 =
π π 2` + 1
`=0
Show that there is no C > 0 independent of k such that
ku0k kL2 (0,π) ≤ CkukH 1 (0,π) = Cπ .

Chapter 7
Linear Hyperbolic Equations
We consider the initial-boundary value problem
utt + Lu = f in Ω × (0, T ) , (7.1a)

u = u0 on Ω × {t = 0} , (7.1b)
ut = u1 on Ω × {t = 0} , (7.1c)
with either the Dirichlet boundary condition
u=0 on ∂Ω × (0, T ) (7.1d)
or the Neumann boundary condition

∂u
aij Ni = g on ∂Ω × (0, T ) , (7.1e)
∂xj
where, as for the parabolic case, L is a time-dependent elliptic operator
defined by
∂ ij ∂u ∂u
(Lu)(x, t) = − a (x, t) (x, t) + bi (x, t) (x, t) (7.2)
∂xi ∂xj ∂xi
+ c(x, t)u(x, t)
satisfying the uniform hyperbolicity condition that for some constant λ 0 > 0,
aij (x, t)ξj ξi ≥ λ 0 |ξ|2 for all (x, t) ∈ Ω × (0, T ), ξ ∈ Rn .
7.1 Weak solutions
We assume that the matrix A = [aij ] is symmetric; that is, aij = aji , and
aij , aij i 2 2 2
t , b , c ∈ C(Ω × [0, T ]), f ∈ L (0, T ; L (Ω)), g ∈ L (0, T ; H
−0.5 (∂Ω))
270
7.1. WEAK SOLUTIONS 271
with gt ∈ L2 (0, T ; H −0.5 (∂Ω)), u0 ∈ H 1 (Ω) and u1 ∈ L2 (Ω). We remark that

if aij , bi , c are time-independent, then only aij , bi , c ∈ L∞ (Ω) are required.
Definition 7.1 (Weak solutions with Dirichlet boundary conditions). A

function u ∈ L∞ (0, T ; H01 (Ω)) with time-derivatives ut ∈ L∞ (0, T ; L2 (Ω))
and utt ∈ L2 (0, T ; H −1 (Ω)) is said to be a weak solution of the hyperbolic
initial-boundary value problem (7.1a,b,c,d) provided that

utt (t), ϕ H 1 (Ω) + B u(t), ϕ = f (t), ϕ L2 (Ω)
0
(7.3)
∀ ϕ ∈ H01 (Ω), a.e. t ∈ (0, T ) ,
and
u = u0 , ut = u1 on Ω × {t = 0} , (7.4)

where B u(t), ϕ is defined by
∂u ∂ϕ
Z h
aij (x, t)

B u(t), ϕ ≡ (x, t) (x)
Ω ∂xj ∂xi
(7.5)
i ∂u i
+ b (x, t) (x, t)ϕ(x) + c(x, t)u(x, t)ϕ(x) dx .
∂xi
Equation (7.3) is called the variational formulation of the hyperbolic problem
(7.1a,b,c,d).
Definition 7.2 (Weak solutions with Neumann boundary conditions). A

function u ∈ L∞ (0, T ; H 1 (Ω)) with time-derivatives ut ∈ L∞ (0, T ; L2 (Ω))
and utt ∈ L2 (0, T ; H 1 (Ω)0 ) is said to be a weak solution of the hyperbolic
initial-boundary value problem (7.1a,b,c,e) provided that

utt (t), ϕ H 1 (Ω) + B u(t), ϕ = f (t), ϕ L2 (Ω) + g(t), ϕ H 0.5 (∂Ω)
0
(7.6)
∀ ϕ ∈ H01 (Ω), a.e. t ∈ (0, T ) ,

and (7.4) holds. Here again B u(t), ϕ is defined by (7.5). Equation (7.6)
is called the variational formulation of the hyperbolic problem (7.1a,b,c,e).
Remark 7.3. In order to establish the existence of the weak solution to the
hyperbolic equation, we already have to assume the symmetry of A, while
the same condition is used to prove the regularity theory of the parabolic
problem.
272 CHAPTER 7. Linear Hyperbolic Equations
Remark 7.4. Let u be a weak solution to the hyperbolic problem (7.1a,b,c,d).

By definition u ∈ C([0, T ]; L2 (Ω)) and ut ∈ C([0, T ]; H −1 (Ω)); thus we can
expect that
lim ku(t) − u0 kL2 (Ω) = lim kut (t) − u1 kH −1 (Ω) = 0. (7.7)

t→0 t→0
Moreover, by Lemma 6.8, u ∈ C([0, T ]; H01 (Ω)w ) and ut ∈ C([0, T ]; L2 (Ω)w );

that is,
lim hu(t) − u(t0 ), ϕiH01 (Ω) = 0 ∀ ϕ ∈ H −1 (Ω) , t0 ∈ [0, T ] ,

t→t0
∀ ϕ ∈ L2 (Ω) , t0 ∈ [0, T ] ,

lim ut (t) − ut (t0 ), ϕ L2 (Ω) = 0
t→t0
so in addition to (7.7), we also have
∀ ϕ ∈ H −1 (Ω),

lim u(t) − u0 , ϕ H 1 (Ω) = 0
t→0
0
lim ut (t) − u1 , ϕ L2 (Ω) = 0 ∀ ϕ ∈ L2 (Ω).
t→0
Similar arguments apply to the hyperbolic problem (7.1a,b,c,e). Therefore,

(7.4) in the definition of weak solutions is understood as
lim ku(t) − u0 kL2 (Ω) = lim kut (t) − u1 kX 0 = 0, (7.8)

t→0 t→0
and
lim hu(t) − u0 , ϕiX = 0 ∀ ϕ ∈ X 0,
t→t0
(7.9)
∀ ϕ ∈ L2 (Ω),

lim ut (t) − u1 , ϕ L2 (Ω) = 0
t→t0
where X = H01 (Ω) for Definition 7.1 or X = H 1 (Ω) for Definition 7.2.
Remark 7.5. As for the parabolic case, the argument leading (6.14) to
(6.11) implies that the variational formulation (7.3) is equivalent to
Z Th

i
utt (t), ϕ(t) H 1 (Ω)+ B u(t), ϕ(t) dt
0
0
(7.10)
Z T
2 1

= f (t), ϕ(t) L2 (Ω)
dt ∀ ϕ ∈ L (0, T ; H (Ω)),
0
and the variational formulation (7.6) is equivalent to

Z Th

i
utt (t), ϕ(t) H 1 (Ω)+ B u(t), ϕ(t) dt (7.11)
0
Z Th i
f (t), ϕ(t) L2 (Ω)+ g(t), ϕ(t) L2 (∂Ω) dt ∀ ϕ ∈ L2 (0, T ; H 1 (Ω)).

=
0
7.1. WEAK SOLUTIONS 273
In the following, we focus only on the discussion of the hyperbolic equa-

tion with the Neumann boundary condition, while similar argument can be
made for the hyperbolic equation with the Dirichlet boundary condition.
We start with the uniqueness of the weak solution.
7.1.1 Uniqueness of the weak solution
Assume that u1 and u2 are two weak solutions. Let w = u1 − u2 . Then by

(7.11) w satisfies
Z Th
i
∀ v ∈ L2 (0, T ; H 1 (Ω)), (7.12a)

wtt (t), v(t) H 1 (Ω)+ B w(t), v(t) dt = 0
0
w = wt = 0 at t = 0. (7.12b)
Since w ∈ L∞ (0, T ; H 1 (Ω)), the function v defined by

 Z s
 w(τ )dτ if 0 ≤ t ≤ s ,
v(t) = t
0 if s < t ≤ T .

belongs to L2 (0, T ; H 1 (Ω)); thus v can be used as a test function in (7.12a).

Integrating by parts in time,

t=s Z sh i
wt (t), v (t) H 1 (Ω) − wt (t), vt (t) L2 (Ω) − B w(t), v (t) dt = 0 .

t=0 0
(7.13)
Therefore, by Lemma 6.9 and (7.12b), v(s) = 0, vt = −w in [0, s],

t=s Z s 1 s d
Z
kw(t)k2L2 (Ω) dt.

wt (t), v(t) H 1 (Ω) − wt (t), vt (t) L2 (Ω) dt =

t=0 0 2 0 dt
By the symmetry of A = [aij ],
ij ∂vt ∂v 1d ij ∂v ∂v 1 ∂v ∂v
Z Z Z
a dx = a dx − aij
t dx .
Ω ∂xj ∂xi 2 dt Ω ∂xj ∂xi 2 Ω ∂xj ∂xi
Moreover, integrating by parts in t, by v(s) = 0 we obtain that

sZ sZ
i ∂vt ∂2v
Z Z
b vdxdt = bi v dxdt
0 Ω ∂xi 0 Ω ∂xi ∂t
Z sZ
∂v ∂v
Z
= bi (0)v(0) (0) dx − (bi v)t dx .
Ω ∂xi 0 Ω ∂xi
As a consequence, (7.13) implies that

s
1 dh ∂v ∂v
Z Z i
2
kw(t)kL2 (Ω) − aij (t) (t) (t) dx dt
0 2 dt Ω ∂xj ∂xi
Z sh Z
1 ∂v ∂v ∂v
Z Z i
=− aij dx + (bi
v)t dx + cwvdx dt
0 2 Ω t ∂xj ∂xi Ω ∂xi Ω
∂v(0)
Z
+ bi (0) v(0) dx .
Ω ∂xi
Again by w(0) = 0 and v(s) = 0, as well as the uniform hyperbolicity

assumption,
Z sh
kw(s)k2L2 (Ω) + λ 0 kDv(0)k2L2 (Ω) ≤ kat (t)kL∞ (Ω) kDv(t)k2L2 (Ω)
0

+ kbt (t)kL∞ (Ω) kv(t)kL2 (Ω) + kb(t)kL∞ (Ω) kw(t)kL2 (Ω) kDv(t)kL2 (Ω)
i
+ kc(t)kL∞ (Ω) kw(t)kL2 (Ω) kv(t)kL2 (Ω) dt
λ0 1
+ kDv(0)k2L2 (Ω) + kb(0)k2L∞ (Ω) kv(0)k2L2 (Ω)
2 2λ 0
Z sh
≤ kat (t)kL∞ (Ω) + kbt (t)kL∞ (Ω) + kb(t)kL∞ (Ω) + kc(t)kL∞ (Ω)
0
s ih i λ0
kb(0)k2L∞ (Ω) kv(t)k2H 1 (Ω) + kw(t)k2L2 (Ω) dt +
+ kDv(0)k2L2 (Ω)
2λ 0 2
Z sh
≤ (1 + s) kat (t)kL∞ (Ω) + kbt (t)kL∞ (Ω) + kb(t)kL∞ (Ω) + kc(t)kL∞ (Ω)
0
kb(0)k2L∞ (Ω) ih i λ0
+ kDv(t)k2L2 (Ω) + kw(t)k2L2 (Ω) dt + kDv(0)k2L2 (Ω) .
2λ 0 2
Z t
Let W (t) = w(τ )dτ . Then the inequality above implies that
0
2kw(s)k2L2 (Ω) + λ 0 kDW (s)k2L2 (Ω)

Z sh i
≤ C(1+ s) kDW (t)− DW (s)k2L2 (Ω) + kw(t)k2L2 (Ω) dt
Z0 s h i
≤ C(1+ s) kDW (t)k2L2 (Ω) + kw(t)k2L2 (Ω) dt+ C(s + s2 )kDW (s)k2L2 (Ω)
0
or
2kw(s)k2L2 (Ω) + λ 0 − C(s + s2 ) kDW (s)k2L2 (Ω)

Z sh i
≤ C(1 + s) kDW (t)k2L2 (Ω) + kw(t)k2L2 (Ω) dt .
0
7.2. THE METHOD OF PARABOLIC REGULARIZATION - THE SIMPLE CASE275
For T1 > 0 small enough so that λ 0 − C(T1 + T12 ) ≥ λ 0 /2, the Gronwall
inequality implies that for all t ∈ [0, T1 ],
h i
kw(t)k2L2 (Ω) + kDW (t)k2L2 (Ω) ≤ C kw(0)k2L2 (Ω) + kDW (0)k2L2 (Ω) = 0 ;
hence w(t) = 0 or u1 (t) = u2 (t) in [0, T1 ]. In other words, the solution to

(7.1a,b,c,e) is unique in the time interval [0, T1 ]. The same argument applies
to the interval [T1 , 2T1 ], [2T1 , 3T1 ], etc., so the solution is unique in the time
interval [0, T ]. Therefore, we establish the following
Theorem 7.6. There exists at most one weak solution to the hyperbolic
initial-boundary value problem (7.1a,b,c,d) or (7.1a,b,c,e).
7.2 The method of parabolic regularization - the

simple case
Given the uniqueness of the weak solution, we only focus on the existence
and the regularity theory in the rest of the discussion. To illustrate the idea,
we start with the wave equation; that is, aij = δij and bi = c = 0, and will
return to the general case later.
7.2.1 The existence of the weak solution
We first study the existence of a weak solution solution to the hyperbolic

initial-boundary value problem
utt − ∆u = f in Ω × (0, T ) , (7.14a)

∂u
=g on ∂Ω × (0, T ) , (7.14b)
∂N
u = u0 , ut = u1 on Ω × {t = 0} . (7.14c)
We would like to prove the following
Theorem 7.7. There exists one and only one weak solution to the wave
equation (7.14).
It suffices to show the existence of a weak solution to (7.14). To prove

the existence, instead of considering (7.14) directly, we consider the following
parabolic κ-problem
uκtt − κ∆uκt − ∆uκ = f in Ω × (0, T ) , (7.15a)

∂uκ ∂uκ
κ t + =g on ∂Ω × (0, T ) , (7.15b)
∂N ∂N
uκ = u0 , uκt = u1 on Ω × {t = 0} , (7.15c)
where κ > 0 is artificial viscosity introduced into the system, and look for the
(weak) limit of uκ as κ → 0. The superscript κ indicates that the solution
uκ depends on the parameter κ. We remark that by letting w = uκt , (7.15a)
is a parabolic equation of w (if ∆uκ is treated as a forcing function), and
(7.15) is called the zeroth-order parabolic regularization of (7.14).
One of the strategy of solving (7.15) for all fixed κ > 0 is through a
Z t
fixed-point argument: given w̃ ∈ L2 (0, T ; H 1 (Ω)), let ũ = u0 + w̃(s) ds,
0
and solve
wt − κ∆w = f + ∆ũ in Ω × (0, T ) , (7.16a)

∂w ∂ ũ
κ =g− on ∂Ω × (0, T ) , (7.16b)
∂N ∂N
w = u1 on Ω × {t = 0} . (7.16c)
By the discussion in Section 6.3 the weak solution to (7.16) exists; thus we
establish a map Φ : L2 (0, T ; H 1 (Ω)) → L2 (0, T ; H 1 (Ω)) by Φ(w̃) = w, and
a fixed-point of Φ is a weak solution to (7.15).
The existence of a fixed-point relies on some type of fixed-point theorem.
Here we will employee the famous contraction mapping theorem.
Lemma 7.8 (Contraction mapping theorem). Let X be a complete metric

space, and Ψ : X → X be a contraction mapping on X; that is, for some
constant 0 < C < 1,

dX Ψ(x), Ψ(y) ≤ CdX (x, y) ∀ x, y ∈ X ,
where dX denotes the metric on X. Then Φ has a unique fixed-point in X.
We show that Φ is a contraction mapping on L2 (0, T ; H 1 (Ω)) provided

that T > 0 is small enough. Now given w̃1 and w̃2 in L2 (0, T ; H 1 (Ω)), let
§7.2 The method of parabolic regularization - the simple case 277
Z t
ũi = u0 + w̃i (s) ds, δ ũ = ũ1 − ũ2 , wi = Φ(w̃i ), and δ w̃ = w̃1 − w̃2 . Then
0
δw = w1 − w2 satisfies δw(0) = 0, and

(δw)t , ϕ H 1 (Ω) + κ(Dδw, Dϕ)L2 (Ω) = Dδ ũ, Dϕ L2 (Ω)
(7.17)
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ) .
Since δw ∈ L2 (0, T ; H 1 (Ω)), δw(t) ∈ H 1 (Ω) for almost all t ∈ (0, T ); thus
can be used as a test function in (7.17). By doing so, we find that
1d
kδwk2L2 (Ω) + κkDδwk2L2 (Ω) = (Dδ ũ, Dδw)L2 (Ω)
2 dt
1 κ
≤ kDδ ũk2L2 (Ω) + kDδwk2L2 (Ω) .
2κ 2
Therefore, by δw(0) = 0,
t t
1
Z Z
kδw(t)k2L2 (Ω) + κ kDδw(s)k2L2 (Ω) ds ≤ kDδ ũ(s)k2L2 (Ω) ds
0 κ 0
Z t
Now, since δ ũ = δ w̃(s) ds,
0
Z t √ hZ t i1
2
kDδ ũkL2 (Ω) ≤ kDδ w̃(s)kL2 (Ω) ds ≤ t kDδ w̃(s)k2L2 (Ω) ds .
0 0
As a consequence,
Z t
kδw(t)k2L2 (Ω)
+κ kDδw(s)k2L2 (Ω) ds
0
1 t
Z s
t2 t
Z Z
0 2 0
≤ s kDδ w̃(s )kL2 (Ω) ds ds ≤ kDδ w̃(s)k2L2 (Ω) ds .
κ 0 0 2κ 0
Moreover, integrating in time, the energy estimate above implies that

t t
t3
Z Z
kδw(s)k2L2 (Ω) ds ≤ kDδ w̃(s)k2L2 (Ω) ds ;
0 6κ 0
thus
t h t3 t
t2 i
Z Z
kδw(s)k2H 1 (Ω) ≤ + 2 kDδ w̃(s)k2L2 (Ω) ds .
0 6κ 2κ 0
Note that the inequality above is the same as

r
t3 t2
kΦ(w̃1 ) − Φ(w̃2 )kL2 (0,T ;H 1 (Ω)) ≤ + 2 kw̃1 − w̃2 kL2 (0,T ;H 1 (Ω)) ; (7.18)
6κ 2κ
hence the contraction mapping theorem implies that there is a Tκ > 0 such
that the unique weak solution to (7.15) exists in the time interval [0, Tκ ].
Having established the existence of the weak solution uκ to (7.15), the
next step is to passing κ to the limit which requires a κ-independent up-
per bounded for uκ ∈ L∞ (0, Tκ ; H 1 (Ω)), uκt ∈ L∞ (0, Tκ ; L2 (Ω)) and uκtt ∈
L2 (0, Tκ ; H 1 (Ω)0 ). Moreover, from the process of constructing the weak so-
lution, the time of existence Tκ appears to approach 0 as κ → 0 (so that
the contraction constant is strictly smaller than 1); thus in addition to a
uniform upper bound of uκ , we also need to prove that the time of existence
can be made independent of κ, then pass κ → 0 to obtain the weak solution
to (7.14).
κ-independent estimates
Z t
κ
The unique fixed point w satisfies u (t) = u0 + w(s)ds or uκt = w. There-
0
fore, we may rewrite the variational formulation of (7.16) as

κ
utt (t), ϕ H 1 (Ω) + κ Duκt (t), Dϕ L2 (Ω) + Duκ (t), Dϕ L2 (Ω)

= f (t), ϕ L2 (Ω) + g, ϕ H 0.5 (∂Ω) ∀ ϕ ∈ H 1 (Ω) , a.e. t ∈ (0, T ).

(7.19)
Since uκt = w ∈ L2 (0, Tκ ; H 1 (Ω)), it can be used as a test function, and we
obtain that
1 dh κ 2 i
kut kL2 (Ω) + kDuκ k2L2 (Ω) + κkDuκt k2L2 (Ω)
2 dt
1 1
≤ kf k2L2 (Ω) + kuκt k2L2 (Ω) + hg, uκt iH 0.5 (∂Ω) .
2 2
Similar to (6.27),
Z t s=t Z t
hg, uκt iH 0.5 (∂Ω) ds κ
hgt , uκ iH 0.5 (∂Ω) ds

= g(s), u (s) H 0.5 (∂Ω) −

0 s=0 0
h i
≤ C kg(0)k2H −0.5 (∂Ω) + ku0 k2H 1 (Ω) + C kg(t)k2H −0.5 (∂Ω) (7.20)
Z t
+ kuκ (t)k2H 1 (Ω) + Ckgt k2L2 (0,T ;H −0.5 (∂Ω)) + C kuκ (s)k2H 1 (Ω) ds.
0
Therefore, integrating the inequality above in time over the time interval
(0, t) and choosing > 0 small enough, we find that
Z t
kuκt (t)k2L2 (Ω) + kDu κ
(t)k2L2 (Ω) + κ kDuκt (s)k2L2 (Ω) ds
0
h
≤C kuκt (0)k2L2 (Ω) + kDu κ
kf k2L2 (0,T ;L2 (Ω)) + kgk2L2 (0,T ;H −0.5 (∂Ω))
(0)k2L2 (Ω) +
i Z th i
2
+ kgt kL2 (0,T ;H −0.5 (∂Ω)) + C kuκt (s)k2L2 (Ω) + kDuκ (s)k2L2 (Ω) ds ;
0
thus the Gronwall inequality and (6.26) imply that

Z t
kuκt (t)k2L2 (Ω) + kDuκ (t)k2L2 (Ω) + κ kDuκt (s)k2L2 (Ω) ds
0
h
≤ C ku0 kH 1 (Ω) + ku1 kL2 (Ω) + kf k2L2 (0,T ;L2 (Ω))
2 2
(7.21)
i
+ kgk2L2 (0,T ;H −0.5 (∂Ω)) + kgt k2L2 (0,T ;H −0.5 (∂Ω)) eCt ;
thus uκ is uniformly bounded in L∞ (0, T ; H 1 (Ω)), uκt is uniformly bounded

√
in L∞ (0, T ; L2 (Ω)), and κDuκt is uniformly bounded in L2 (0, Tκ ; L2 (Ω)).
We also need to establish a uniform bound for uκtt in L2 (0, T ; H 1 (Ω)0 ).
By the variational formulation (7.19) we find that
kuκtt kH 1 (Ω)0 = sup huκtt , ϕiH 1 (Ω) ≤ kf kL2 (Ω) + κkDuκt kL2 (Ω) + kDuκ kL2 (Ω) ;
kϕkH 1 (Ω) =1
thus (7.21) implies that

Z t h
kuκtt (s)k2H 1 (Ω)0 ds ≤ C ku0 k2H 1 (Ω) + ku1 k2L2 (Ω) + kf k2L2 (0,T ;L2 (Ω)) (7.22)
0
i
+ kgk2L2 (0,T ;H 0.5 (∂Ω)) + kgt k2L2 (0,T ;H −0.5 (∂Ω)) .
The combination of (7.21) and (7.22) then gives

h i
max kuκt (t)kL2 (Ω) + kuκ (t)kH 1 (Ω) + kuκtt (t)kL2 (0,T ;H 1 (Ω)0 )
t∈[0,T ]
h
≤ C ku0 kH 1 (Ω) + ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) (7.23)
i
+ kgkL2 (0,T ;H 0.5 (∂Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ).

A κ-independent time of existence
By the contraction mapping theorem, the solution uκ exists in a time interval

of length at least Tκ . If the solution uκ fails to exist, beyond certain T > Tκ
the function uκ will fail to belong to L2 (0, T ; H 1 (Ω)) which is the same as
saying that kuκ kL2 (0,T ;H 1 (Ω)) = ∞. However, we note that (7.21) holds as
long as the solution exists, so it is impossible that uκ blows up in finite time.
Therefore, the time of existence Tκ can be make independent of κ, and we
set Tκ = T . Moreover, the solution exists uniquely in the time interval [0, T ].
The limit as κ approaches zero

√
By the uniform bound of κDuκt in L2 (0, T ; L2 (Ω)), for all ϕ ∈ H 1 (Ω) we
have
Z T √
Z T √
κ (Duκt , Dϕ)L2 (Ω) dt = κ ( κDuκt , Dϕ)L2 (Ω) dt → 0 as κ → 0.
0 0
Therefore, integrating (7.19) over the time interval (0, T ) and passing κ to
the limit, we obtain that
Z Th i Z T
f (t), ϕ L2 (Ω) dt ∀ ϕ ∈ H 1 (Ω),

hutt (t), ϕiH 1 (Ω) + Du(t), Dϕ L2 (Ω) dt =
0 0
here u is the weak limit of uκ in 2 1
L (0, T ; H (Ω)). This implies that u satisfies
the variational formulation (7.6). Moreover, (7.23) also implies that uκtt is
uniformly bounded in L2 (0, T ; H 1 (Ω)0 ), while Duκ and Duκ are uniformly
bounded in Lp (0, T ; L2 (Ω)). Therefore, the weak limit u satisfies
h
kukLp (0,T ;H 1 (Ω)) + kut kLp (0,T ;L2 (Ω)) + kutt kL2 (0,T ;H 1 (Ω)0 ) ≤ C ku0 kH 1 (Ω)
i
+ ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H 0.5 (∂Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ), independent of p. Since lim khkLp (0,T ) =

p→∞
khkL∞ (0,T ) , we conclude that u satisfies
h
kukL∞ (0,T ;H 1 (Ω)) + kut kL∞ (0,T ;L2 (Ω)) + kutt kL2 (0,T ;H 1 (Ω)0 ) ≤ C ku0 kH 1 (Ω)
i
+ ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) + kgkL2 (0,T ;H 0.5 (∂Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω)) ;
so u ∈ L∞ (0, T ; H 1 (Ω)) with time derivatives ut ∈ L∞ (0, T ; L2 (Ω)) and

utt ∈ L2 (0, T ; H 1 (Ω)0 ). Finally,
Z t Z t
uκ (t) = u0 + w(s) ds and w(t) = u1 + uκtt (s) ds
0 0
imply that u(0) = u0 and ut (0) = u1 . Therefore, u is the unique weak

solution to (7.14).
7.2.2 Regularity theory with smooth data
Before proceeding, we introduce the following spaces.
1. the space of solutions: the space consists of space-time dependent

functions so that one time derivative scales like one space derivative.
To be more concrete, for k ≥ 2 and k ∈ N,
n o
W (T ; Ω) ≡ u ∈ L (0, T ; H (Ω)) ∂tj u ∈ L∞ (0, T ; H k−j (Ω)) if j ≤ k
k ∞ k
k
k∂tj ukL∞ (0,T ;H k−j (Ω)) ;
X
equipped with norm kukW k (T ;Ω) =
j=0
2. the space of interior forcing functions: for k ≥ 2 and k ∈ N,

n
FΩk−2 (T ) ≡ f ∈ C([0, T ]; H k−2 (Ω)) ∂tk−1f ∈ L2 (0, T ; L2 (Ω)),

o
∂tj f ∈ C([0, T ]; H k−j−2 (Ω)) if j ≤ k − 2
equipped with norm
k−2
∂tj kf kC([0,T ];H k−j−2 (Ω)) + k∂tk−1f kL2 (0,T ;L2 (Ω)) ;
X
kf kW k−2 (T ) =
Ω
j=0
3. the space of forcing functions on the boundary: for k ≥ 2 and k ∈ N,

n
k−1.5
G∂Ω (T ) ≡ g ∈ C([0, T ]; H k−1.5 (∂Ω)) ∂tk g ∈ L2 (0, T ; H −0.5 (∂Ω)),

o
∂tj g ∈ C([0, T ]; H k−j−0.5 (∂Ω)) if j ≤ k − 1
equipped with norm
k
k∂tj gkC([0,T ];H k−j−1.5 (∂Ω)) + k∂tkgkL2 (0,T ;H −0.5 (∂Ω)) .
X
kgkG k−1.5 (T ) =
∂Ω
j=0
In order to show that the weak solution u to (7.14) possesses higher reg-
ularity, say u ∈ L2 (0, T ; H m+1 (Ω)), we do not look at the zero-th parabolic
regularization of (7.14), but instead start from the m-times time-differentiated

version of the problem of (7.15)
wm = ∂tm w = ∂tm+1 uκ in Ω × (0, T ) , (7.24a)

wmt − κ∆wm − ∆wm−1 = ∂tm f in Ω × (0, T ) , (7.24b)
∂wm ∂wm−1
κ + = ∂tm g on ∂Ω × (0, T ) , (7.24c)
∂N ∂N
wm = um+1 , wm−1 = um on Ω × {t = 0} , (7.24d)
along with its variational formulation

wmt (t), ϕ H 1 (Ω) + κ Dwm (t), ϕ L2 (Ω) + Dwm−1 (t), ϕ L2 (Ω) (7.25)
= ∂tm f (t), ϕ L2 (Ω) + ∂tm g(t), ϕ H 0.5 (∂Ω) ∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ) ,

where in (7.24d) um and um+1 are given initial data whose precise forms
will be specified later.
The argument of showing that (7.15) has a unique weak solution can be
used to show that (7.24) has a unique weak solution wm ∈ L2 (0, T ; H 1 (Ω))
with wmt ∈ L2 (0, T ; H 1 (Ω)0 ) satisfying (7.25). Integrating (7.25) in time
over the time interval (0, t), we find that
Z t Z t
wm (t), ϕ L2 (Ω)+ κ D wm (s)ds, Dϕ 2 + D wm−1 (s)ds, Dϕ 2
0 L (Ω) 0 L (Ω)

= ∂tm−1f (t)−(∂tm−1f )(0)+ um+1 , ϕ 2 + ∂tm−1g(t)−(∂tm−1g)(0), ϕ 2 ,
L (Ω) L (∂Ω)
Z t
and letting wk = uk+1 + wk+1 (s)ds for k = 1, · · · , m further implies that
0

wm−1 (t), ϕ L2 (Ω) + κ Dwm−1 , Dϕ L2 (Ω) + Dwm−2 , Dϕ L2 (Ω)
= ∂tm−1f (t) − (∂tm−1f )(0) + um+1 − κ∆um − ∆um−1 , ϕ L2 (Ω) (7.26)

∂um ∂um−1
+ ∂tm−1g(t) − (∂tm−1g)(0) + κ + ,ϕ 2 .
∂N ∂N L (∂Ω)
In other words, wm−1 is the weak solution to
∂t wm−1 − κ∆wm−1 − ∆wm−2 = Fm−1 in Ω × (0, T ) ,

∂wm ∂wm−1
κ + = Gm−1 on ∂Ω × (0, T ) ,
∂N ∂N
wm−1 = um , wm−2 = um−1 on Ω × {t = 0} ,
where
Fm−1 = ∂tm−1 f (t) − (∂tm−1 f )(0) + um+1 − κ∆um − ∆um−1 ,

∂um ∂um−1
Gm−1 = ∂tm−1 g(t) − (∂tm−1 g)(0) + κ + .
∂N ∂N
Integrating (7.26) in time over the time interval (0, t) for another (m − 1)
times, we find that w is the weak solution to
wt − κ∆w− ∆uκ = f + δf in Ω × (0, T ) , (7.27a)

∂w ∂uκ
κ + = g + δg on ∂Ω × (0, T ) , (7.27b)
∂N ∂N
w = u1 , uκ = u0 on Ω × {t = 0}. (7.27c)
where
m−1
X t` ∂`f

δf = u`+2 − (0) − κ∆u`+1 − ∆u` , (7.28a)
`! ∂t`
`=0
m−1
X t` t` ∂u`+1 ∂u` ∂ ` g
δg = κ + − ` (0) . (7.28b)
`! `! ∂N ∂N ∂t
`=0
For simplicity, we first assume that the initial data u0 , u1 , the forcing func-
tion f , g are smooth, and satisfy the compatibility conditions
∂u` ∂`g
= ` (0) on ∂Ω, ` = 0, 1, · · · , m − 1,
∂N ∂t
∂ `−1
where u`+1 = (f + ∆u + κ∆ut ) for ` = 1, · · · , m. Note that then
∂t`−1 t=0
δf = 0, and
m−1
X t` ∂u`+1
δg = κ . (7.28b0 )
`! ∂N
`=0
As discussed above, in order to show that the weak solution u to (7.14)

belonging to L2 (0, T ; H m+1 (Ω)), instead of studying (7.15) we focus on
wt − κ∆w = f + ∆uκ in Ω × (0, T ) , (7.29a)

∂w ∂uκ
κ =g− + δg on ∂Ω × (0, T ) , (7.29b)
∂N ∂N
w = u1 on Ω × {t = 0} , (7.29c)
Z t
uκ = u0 + w(s) ds in × Ω × (0, T ) . (7.29d)
0
We emphasize again that the boundary condition (7.29b) is designed in

such a way that the compatibility conditions for the parabolic equation
of w are valid up to m-th order. By the smoothness assumptions on the
initial and boundary data, using the fixed-point argument we obtain that
2m (T ; Ω). The key then is to show that ∂ ` u ∈
the unique solution w ∈ VN t
L2 (0, T ; H m+1−` (Ω)) is uniformly bounded for ` = 0, 1, · · · , m + 1. In the
following, we take m = 1 and m = 2 as examples to illustrate how κ-
independent estimates are obtained.
κ-independent estimates, the case m = 1
We study the time-differentiated version of the problem of (7.29):
wtt − κ∆wt − ∆w = ft in Ω × (0, T ) , (7.30a)

∂wt ∂w
κ + = gt on ∂Ω × (0, T ) , (7.30b)
∂N ∂N
wt = f (0) + ∆u0 + κ∆u1 on Ω × {t = 0} . (7.30c)
We note that by our smoothness assumptions of the forcing function, bound-

ary and initial data, Theorem 6.29 implies that the solution wt to (7.30)
2 (T ; Ω).
exists and belongs to VN
For the uniform bound of ∂t` uκ in L2 (0, T ; H 2−` (Ω)) for ` = 1, 2, we test
(7.30a) against wt and find that
1 dh i
kwt k2L2 (Ω) + kDwk2L2 (Ω) + κkDwt k2L2 (Ω)
2 dt
1 1
≤ kft k2L2 (Ω) + kwt k2L2 (Ω) + hgt , wt iH 0.5 (∂Ω) .
2 2
Integrating the inequality above in time over the time interval (0, t), by
estimates similar to (7.20) we find that
Z t
2 2
kwt (t)kL2 (Ω)+ kDw(t)kL2 (Ω)+ κ kDwt (s)k2L2 (Ω) ds
0
h
≤ C kf (0)kL2 (Ω)+ ku0 kH 2 (Ω)+ ku1 k2H 1 (Ω)+ kft k2L2 (0,T ;L2 (Ω))
2 2
i
+ kgt k2L2 (0,T ;H −0.5 (∂Ω))+ kgtt k2L2 (0,T ;H −0.5 (∂Ω))
h
≤ C ku0 k2H 2 (Ω)+ ku1 k2H 1 (Ω)+ kf k2L2 (0,T ;L2 (Ω))+ kft k2L2 (0,T ;L2 (Ω))
(7.31)
i
+ kgt k2L2 (0,T ;H −0.5 (∂Ω))+ kgtt k2L2 (0,T ;H −0.5 (∂Ω)) .
√
Therefore, we establish the uniform boundedness of uκtt , uκt , and κuκtt in
L∞ (0, T ; L2 (Ω)), L∞ (0, T ; H 1 (Ω)), and L2 (0, T ; H 1 (Ω)), respectively.
The last piece of the puzzle is to obtain a uniform bound for uκ ∈
L2 (0, T ; H 2 (Ω)). This relies on an elliptic-type estimate, and we first focus
on the following
Lemma 7.9. Let u0 ∈ H m (Ω) and ut ∈ L2 (0, T ; H m (Ω)) satisfies
κut + u = F
for some F ∈ C([0, T ]; H m (Ω)). Then
max ku(t)kH m (Ω) ≤ ku0 kH m (Ω) + max kF (t)kH m (Ω) . (7.32)

t∈[0,T ] t∈[0,T ]
Proof. For almost all t ∈ [0, T ],
kκut (t) + u(t)k2H m (Ω) = kF (t)k2H m (Ω)
which implies
κ2 kut k2H m (Ω) + 2κ(ut , u)H m (Ω) + kuk2H m (Ω) = kF k2H m (Ω) .
By Lemma 6.9,
d
κ ku(t)k2H m (Ω) + ku(t)k2H m (Ω) ≤ kF (t)k2H m (Ω) . (7.33)
dt
Suppose that
max ku(t)kH m (Ω) = ku(t0 )kH m (Ω) . (7.34)

t∈[0,T ]
If t0 = 0, then max ku(t)kH m (Ω) ≤ ku0 kH m (Ω) . If t0 > 0, we integrate (7.33)

t∈[0,T ]
in time over the time interval (t, t0 ) with t < t0 , and Young’s inequality
implies that
h i Z t0 Z t0
2 2
κ ku(t0 )kH m (Ω) − ku(t)kH m (Ω) + ku(s)k2H m (Ω) ds ≤ kF (s)k2H m (Ω) ds.
t t
Assumption (7.34) then further implies that

Z t0 Z t0
ku(s)k2H m (Ω) ds ≤ kF (s)k2H m (Ω) ds ∀ t ∈ (0, t0 ).
t t
Dividing through both sides by t0 − t and passing t → t0 , the fundamental

theorem of Calculus suggests that
max ku(t)kH m (Ω) = ku(t0 )kH m (Ω) ≤ kF (t0 )kH m (Ω) ≤ max kF (t)kH m (Ω)
t∈[0,T ] t∈[0,T ]
which concludes the lemma.
The following corollary is important for the discussion in the rest of the
chapter.
Corollary 7.10. Let ∂Ω be of class C m+1 , and aij ∈ C m (Ω×[0, T ]) satisfies

aij ξj ξi ≥ λ 0 |ξ|2 for some constant λ 0 > 0. Suppose that w = κut + u is a
weak solution to
∂ ij ∂w
− a =f in Ω × (0, T ) , (7.35a)
∂xi ∂xj
∂w
aij Ni = g on ∂Ω × (0, T ) . (7.35b)
∂xj
for some f and g.
1. If u0 ∈ H m+1 (Ω), f ∈ L2 (0, T ; H m−1 (Ω)) and g ∈ L2 (0, T ; H m−0.5 (∂Ω)),

then u ∈ C([0, T ]; H m+1 (Ω));
2. If in addition f ∈ C([0, T ]; H m−1 (Ω)) and g ∈ C([0, T ]; H m−0.5 (∂Ω)),

then u ∈ C([0, T ]; H m+1 (Ω)) satisfies
h
max ku(t)kH m+1 (Ω) ≤ ku0 kH m+1 (Ω) + C max kf (t)kH m−1 (Ω)
t∈[0,T ] t∈[0,T ]
i
+ kg(t)kH m−0.5 (∂Ω) + ku(t)kL2 (Ω)
(7.36)
for some C = C(kakC m (Ω×[0,T ]) , λ 0 ).
Proof. By the elliptic regularity

h i
kwkH m+1 (Ω) ≤ C kf kH m−1 (Ω) + kgkH m−0.5 (∂Ω) + kwkL2 (Ω) (7.37)
for some C = C(kakC m (Ω×[0,T ]) , λ 0 ). Moreover, solving κut + u = w implies

that
t
1
Z
t−s
− κt
u(t) = e u0 + e− κ w(s) ds ; (7.38)
κ 0
thus u ∈ C([0, T ]; H m+1 (Ω)) and ut = κ−1 (w − u) ∈ L2 (0, T ; H m+1 (Ω)).

(7.36) then follows from Lemma 7.9.
Remark 7.11. For the case m = 1, estimate (7.37) and that u is uni-
formly bounded in H 2 (Ω) imply that κut is uniformly bounded in H 2 (Ω).
√
On the other hand, estimate (7.31) implies that κut is uniformly bounded
√
in H 1 (Ω); thus a multiplication of κ gives extra (uniform) spatial regular-
ity.
Now we turn our attention back to the uniform bound of the solution uκ
of (7.29) in L2 (0, T ; H 2 (Ω)). First of all, by (7.31),
Z t
κ
ku (t)kL2 (Ω) ≤ ku0 kL2 (Ω) + kuκt (s)kL2 (Ω) ds
0
Z th Z s i
≤ ku0 kL2 (Ω) + ku1 kL2 (Ω) + kuκtt (s 0 )kL2 (Ω) ds 0 ds
0 0
h
≤ C ku0 kH 2 (Ω) + ku1 kH 1 (Ω) + kft k2L2 (0,T ;L2 (Ω))
2 2
(7.39)
i
+ kgt k2L2 (0,T ;H −0.5 (∂Ω)) + kgtt k2L2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ). Since uκ satisfies
−κ∆uκt − ∆uκ = f − uκtt in Ω,

∂uκ ∂uκ ∂u1
κ t + =g+κ on ∂Ω ,
∂N ∂N ∂N
by Corollary 7.10,
h
max kuκ (t)kH 2 (Ω) ≤ ku0 kH 2 (Ω) + C max kf (t) − uκtt (t)kL2 (Ω)
t∈[0,T ] t∈[0,T ]
∂u1 i
+ g(t) + κ + kuκ (t)kL2 (Ω) .

∂N H (∂Ω)
0.5
As a consequence, we conclude from (7.31) and (7.39) that
max kuκ (t)kH 2 (Ω)

t∈[0,T ]
h
≤ ku0 kH 2 (Ω) + C max kf (t)kL2 (Ω) + kuκtt (t)kL2 (Ω) + kg(t)kH 0.5 (∂Ω)
t∈[0,T ]
i
+ κku1 kH 2 (Ω) + kuκ (t)kL2 (Ω)
h i
≤ C ku0 kH 2 (Ω)+ ku1 kH 1 (Ω)+ κku1 kH 2 (Ω)+ kf kW 0 (T )+ kgkG 0.5 (T ) .
Ω ∂Ω
(7.40)
This estimate implies that uκ is uniformly bounded in L∞ (0, T ; H 2 (Ω)),

while Duκt and uκtt are uniformly bounded in L∞ (0, T ; L2 (Ω)) by (7.31);
thus the weak limit u of uκ in Lp (0, T ; H 2 (Ω)) satisfies
kukL∞ (0,T ;H 2 (Ω)) + kut kL∞ (0,T ;H 1 (Ω)) + kutt kL∞ (0,T ;L2 (Ω))
h i (7.41)
≤ C ku0 kH 2 (Ω) + ku1 kH 1 (Ω) + kf kW 0 (T ) + kgkG 0.5 (T ) .
Ω ∂Ω
∂u0
Moreover, by the compatibility condition = g(0), uκ satisfies the vari-
∂N
ational formulation
uκtt (t), ϕ L2 (Ω) + κ Duκt (t), Dϕ L2 (Ω) + Duκ (t), Dϕ L2 (Ω)

(7.42)
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ).

= f (t), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω)
Therefore, integrating (7.42) in time over the time interval (0, T ) and pass-
ing κ → 0, we find that there exists a u ∈ L2 (0, T ; H 2 (Ω)) with ut ∈
L2 (0, T ; H 1 (Ω)) and utt ∈ L2 (0, T ; L2 (Ω)) satisfying
Z Th i

utt (s), ϕ L2 (Ω) + Du(s), Dϕ L2 (Ω) ds
0
Z Th i
∀ ϕ ∈ H 1 (Ω)

= f (s), ϕ L2 (Ω) + g(s), ϕ L2 (∂Ω) ds
0
Remark 7.5 implies u satisfies

utt (t), ϕ L2 (Ω) + Du(t), Dϕ L2 (Ω)
∀ ϕ ∈ H 1 (Ω), a.e. t ∈ (0, T ) .

= f (t), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω)
Moreover, taking the limit of

Z t Z t
κ
u (t) = u0 + uκt (s) ds, uκt (t) = u1 + uκtt (s) ds,
0 0
we conclude that u(0) = u0 and ut (0) = u1 . Thus we prove the weak solution
u to (7.14) indeed possesses better regularity and satisfies (7.41), provided
that the initial data and the forcing functions are smooth, as well as the
first-order compatibility condition.
κ-independent estimates, the case m = 2
For the case m = 2, we assume that first-order and second-order compati-

bility conditions
∂u0 ∂u1
= g(0), = gt (0) on ∂Ω.
∂N ∂N
Note that the compatibility conditions imply that
kg(0)kH 1.5 (∂Ω) ≤ Cku0 kH 3 (Ω) , kgt (0)kH 0.5 (∂Ω) ≤ Cku1 kH 2 (Ω) .
In order to obtain uκ ∈ L∞ (0, T ; H 3 (Ω)) with κ-independent estimates, we

need to show that wt ∈ L∞ (0, T ; H 1 (Ω)). It requires the study of
wttt − κ∆wtt − ∆wt = ftt in Ω × (0, T ) , (7.43a)

∂wtt ∂wt
κ + = gtt on ∂Ω × (0, T ) , (7.43b)
∂N ∂N
wt = u2 , wtt = u3 on Ω × {t = 0} , (7.43c)
which is the second time-differentiated version of the problem of
wt − κ∆w = f + ∆uκ in Ω × (0, T ) , (7.44a)

∂w ∂uκ ∂u1 ∂u2
κ + =g+κ + κt on ∂Ω × (0, T ) , (7.44b)
∂N ∂N ∂N ∂N
w = u1 on Ω × {t = 0} , (7.44c)
Z t
κ
u (t) = u0 + w(s) ds, (7.44d)
0
where in (7.43c) and (7.44b), u2 ≡ f (0)+∆u0 +κ∆u1 and u3 = ft (0)+∆u1 +

κ∆u2 . The weak solution wtt to (7.43) belongs to L2 (0, T ; H 1 (Ω)) with time
derivative wttt ∈ L2 (0, T ; H 1 (Ω)0 ), and similar to (7.31), w satisfies
h i Z T
sup kwtt (t)k2L2 (Ω)+ kDwt (t)k2L2 (Ω) +κ kDwtt (s)k2L2 (Ω) ds
t∈[0,T ] 0
h
≤ C ku3 k2L2 (Ω)+ ku2 k2H 1 (Ω)+ kftt k2L2 (0,T ;L2 (Ω))+ kgtt k2L2 (0,T ;H −0.5 (∂Ω))
i
+ kgttt k2L2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ). By the second-order compatibility condition

∂u1
= gt (0) on ∂Ω,
∂N
wt satisfies (the variational formulation of)
wtt − κ∆wt − ∆w = ft in Ω × (0, T ) , (7.45a)

∂wt ∂w ∂u1
κ + = gt + κ on ∂Ω × (0, T ) , (7.45b)
∂N ∂N ∂N
w = u1 , wt = u2 on Ω × {t = 0}; (7.45c)
thus by Corollary 7.10,

h
max kw(t)kH 2 (Ω) ≤ ku1 kH 2 (Ω) + C max kft (t) − wtt (t)kL2 (Ω)
t∈[0,T ] t∈[0,T ]
∂u1 i
+ gt (t) + κ + kw(t)k

2
L (Ω)
∂N H 0.5 (∂Ω)

h
≤ C ku3 kL2 (Ω) + ku2 kH 1 (Ω) + (1 + κ)ku1 kH 2 (Ω)
+ kft kL2 (0,T ;L2 (Ω)) + kftt kL2 (0,T ;L2 (Ω)) + kgt kC([0,T ];H 0.5 (∂Ω))
i
+ kgtt kL2 (0,T ;H −0.5 (∂Ω)) + kgttt kL2 (0,T ;H −0.5 (∂Ω))
for some constant C = C(T ). By the first-order compatibility condition

∂u0
= g(0) on ∂Ω,
∂N
w satisfies (7.44); thus by Corollary 7.10 again,
max kuκ (t)kH 3 (Ω)

t∈[0,T ]
h
≤ ku0 kH 3 (Ω) + C ku3 kL2 (Ω) + ku2 kH 1 (Ω) + κ ku1 kH 3 (Ω) + ku2 kH 3 (Ω)
+ kf kC([0,T ];H 1 (Ω)) + kftt kL2 (0,T ;L2 (Ω)) + kgkC([0,T ];H 1.5 (∂Ω))
i
+ kgtt kL2 (0,T ;H −0.5 (∂Ω)) + kgttt kL2 (0,T ;H −0.5 (∂Ω)) .
Since u2 = f (0) + ∆u0 + κ∆u1 and u3 = ft (0) + ∆u1 + κ∆u2 ,
ku3 kL2 (Ω) + ku2 kH 1 (Ω) + κku2 kH 3 (Ω)

≤ kf (0)kH 1 (Ω) + kft (0)kL2 (Ω) + ku0 kH 3 (Ω) + ku1 kH 2 (Ω)
+ κku1 kH 3 (Ω) + κkf (0)kH 3 (Ω) + κku0 kH 5 (Ω) + κ2 ku1 kH 5 (Ω) ,
from which we conclude that

h i
max kuκ (t)kH 3 (Ω)+ kuκt (t)kH 2 (Ω)+ kuκtt (t)kH 1 (Ω)+ kuκttt (t)kL2 (Ω)
t∈[0,T ]
h
≤ C ku0 kH 3 (Ω)+ ku1 kH 2 (Ω)+ κku0 kH 5 (Ω)+ κku1 kH 3 (Ω) (7.46)
i
+ κkf (0)kH 3 (Ω)+ κ2 ku1 kH 5 (Ω)+ kf kF 1 (T ) + kgkG 1.5 (T )
Ω ∂Ω
for some constant C = C(T ). (7.46) further implies that the weak limit u
of uκ satisfies
h i
kukW 3 (T ;Ω) ≤ C ku0 kH 3 (Ω) + ku1 kH 2 (Ω) + kf kW 1 (T ) + kgkG 1.5 (T ) . (7.47)
Ω ∂Ω
We remark here that even though (7.47) does not depend on higher order
norm of u0 , u1 , and f (0), it is required that u0 , u1 , and f (0) to be more
regularity in order to obtain this estimate.
7.2.3 Regularity theory with data possessing limited regu-

larity1
The case m = 1
We first assume that u0 ∈ H 2 (Ω), u1 ∈ H 1 (Ω), f ∈ FΩ0 (T ), and g ∈ G∂Ω

0.5 (T )
as well as the first-order compatibility condition
∂u0
= g(0) on ∂Ω.
∂N
Note that in (7.40), u1 is required to belong to H 2 (Ω) due to the extra

∂u1
term κ that we add into the boundary condition in order to make sure
∂N
the validity of the compatibility condition for the parabolic equation of w.
However, if we can regularize u1 by the parameter κ in such a way that
κku1κ kH 2 (Ω) ≤ Cku1 kH 1 (Ω) and u1κ → u1 as κ → 0, (7.48)
where u1κ denotes some smooth version of u1 , then by considering
wt − κ∆w = f + ∆uκ in Ω × (0, T ) , (7.49a)

∂w ∂uκ ∂u1κ
κ =g− +κ on ∂Ω × (0, T ) , (7.49b)
∂N ∂N ∂N
w = u1κ on Ω × {t = 0} , (7.49c)
Z t
uκ = u0 + w(s) ds in × Ω × (0, T ) , (7.49d)
0
the same procedure implies that uκ satisfies (7.40) with u1 replaced by u1κ ;
thus by (7.48) the weak limit u satisfies (7.41). It remains to show that
(7.48) is valid for some u1κ , and ut (0) = u1 for the weak limit u.
There are several ways of regularizing u1 so that (7.48) holds. The easiest
way should be done by the mollification. Let Ek : H k (Ω) → H k (Rn ) denote
an extension operator satisfying
1. Ek u ∈ H k (Rn ), and Ek u is compactly supported;

1
The reader may skip this section on the first reading
2. Ek u = u in Ω, and
3. Ek u satisfies kEk ukH k (Rn ) ≤ CkukH k (Ω) for all u ∈ H k (Ω).
We define u1κ as ηκ ∗ (E1 u1 ), where ηκ is the usual mollifiers defined in

Section ??. By the property of convolutions,
ku1κ kH k+` (Ω) ≤ Cκ−` ku1 kH k (Ω) ;
thus (7.48) holds.

Finally, by the fundamental theorem of Calculus,
Z t
κ
ut (t) = u1κ + uκtt (s)ds ;
0
thus passing κ to 0, by u1κ → u1 in H 1 (Ω), uκt * ut in L2 (0, T ; H 1 (Ω)), and

uκtt * utt in L2 (0, T ; L2 (Ω)), we conclude that
Z t
ut (t) = u1 + utt (s)ds
0
which implies that ut (0) = u1 .

We also emphasize that in the current discussion we do not modify u0 , so
the first-order compatibility condition of the parabolic equation of w remains
valid. The next discussion will focus on the case that we have to regularize
u0 as well.
The case that m = 2
Similar to the discussion in the previous case, the presence of κku0 kH 5 (Ω) ,
κku1 kH 3 (Ω) , κkf (0)kH 3 (Ω) and κ2 ku1 kH 5 (Ω) in estimate (7.46) implies that
we might need to regularize u0 , u1 , f , as well as g (the reason for the
need of regularizing g can be found in Remark 7.14) in order to conclude
u ∈ L∞ (0, T ; H 3 (Ω)). However, the smooth versions u0κ and u1κ of u0 and
u1 in general do not satisfy
∂u0κ ∂u1κ
= gκ (0), = gκt (0) on ∂Ω
∂N ∂N
which invalidate the compatibility conditions for the parabolic equation
(7.49). As a consequence, integrating the variational formulation of (7.43)
(with u2 and u3 replaced by u2κ and u3κ , while the definition of u2κ and u3κ
will be stated later) in time will lead to the conclusion that wt is the weak
solution to
wtt − κ∆wt − ∆w = fκt + (δfκ )t in Ω × (0, T ) , (7.50a)

∂wt ∂w
κ + = gκt + (δgκ )t on ∂Ω × (0, T ) , (7.50b)
∂N ∂N
w = u1κ , wt = u2κ on Ω × {t = 0} , (7.50c)
and w is the weak solution to
wt − κ∆w − ∆uκ = fκ + δfκ in Ω × (0, T ) , (7.51a)

∂w ∂uκ
κ + = gκ + δgκ on ∂Ω × (0, T ) , (7.51b)
∂N ∂N
w = u1κ on Ω × {t = 0} , (7.51c)
Z t
κ
u (t) = u0κ + w(s) ds, (7.51d)
0
where

δfκ = u2κ − fκ (0) − ∆u0κ − κ∆u1κ + t u3κ − fκt (0) − ∆u1κ − κ∆u2κ ,
∂u0κ ∂u0κ h ∂u
1κ ∂u1κ i
δgκ = κ + − gκ (0) + t κ + − gκt (0) ,
∂N ∂N ∂N ∂N
and u2κ = wt (0) and u3κ = wtt (0) are the initial data for wt and wtt deter-
mined later.
By Corollary 7.10, (7.50) implies that
h
max kw(t)kH 2 (Ω) ≤ ku1κ kH 2 (Ω) + C max kfκt (t) + (δfκ )t − wtt (t)kL2 (Ω)
t∈[0,T ] t∈[0,T ]
i
+ kgκt (t) + (δgκ )t kH 0.5 (∂Ω) + kw(t)kL2 (Ω)
h
≤ C ku3κ kL2 (Ω) + ku2κ kH 1 (Ω) + (1 + κ)ku1κ kH 2 (Ω)
+ ku3κ − fκt (0) − ∆u1κ − κ∆u2κ kL2 (Ω) (7.52)
+ kft kL2 (0,T ;L2 (Ω)) + kftt kL2 (0,T ;L2 (Ω)) + kgt kC([0,T ];H 0.5 (∂Ω))
i
+ kgtt kL2 (0,T ;H −0.5 (∂Ω)) + kgttt kL2 (0,T ;H −0.5 (∂Ω)) ,
here we make use of the estimate

∂u
1κ
− gκt (0) 0.5 ≤ Cku1κ kH 2 (Ω)

∂N

H (∂Ω)
to conclude (7.52). We note that here we do not make the substitution u3κ =
fκt (0) − ∆u1κ − κ∆u2κ yet to get rid of ku3κ − fκt (0) − ∆u1κ − κ∆u2κ kL2 (Ω)
in the estimate. We keep it here for the moment to illustrate one particular
issue that we will encounter later.
Similarly, by Corollary 7.10 again, (7.51) implies that
h
max kuκ (t)kH 3 (Ω) ≤ ku0κ kH 3 (Ω)+ C max kfκ (t)+ δfκ (t)− wt (t)kH 1 (Ω)
t∈[0,T ] t∈[0,T ]
i
+ kgκ (t)+ δgκ (t)kH 1.5 (∂Ω)+ kuκ (t)kL2 (Ω)
h
≤ ku0κ kH 3 (Ω)+ C ku3κ kL2 (Ω)+ ku2κ kH 1 (Ω)+ κ ku1κ kH 3 (Ω)+ ku2κ kH 3 (Ω)
∂u
1κ
+ ku3κ − fκt (0)− ∆u1κ − κ∆u2κ kH 1 (Ω)+ − gκt (0) 1.5

∂N H (∂Ω)
(7.53)
+ kfκ kC([0,T ];H 1 (Ω))+ kfκtt kL2 (0,T ;L2 (Ω))+ kgκ kC([0,T ];H 1.5 (∂Ω))
i
+ kgκtt kL2 (0,T ;H −0.5 (∂Ω))+ kgκttt kL2 (0,T ;H −0.5 (∂Ω)) ,
here we make use of the estimate

∂u
0κ
ku2κ − fκ (0)− ∆u0κ − κ∆u1κ kH 1 (Ω) + − gκ (0) 1.5

h ∂N H (∂Ω) i
≤ C ku0κ kH 3 (Ω) + ku2κ kH 1 (Ω) + κku1κ kH 3 (Ω) + kfκ (0)kH 1 (Ω)
(7.54)
to conclude (7.53). We emphasize that we do not make the substitution
∂u0κ
u2κ = fκ (0) + ∆u0κ + κ∆u1κ and that = gκ (0) to conclude (7.54).
∂N
To obtain a κ-independent estimate for uκ ∈ L∞ (0, T ; H 3 (Ω)), we need
1. u3κ − fκt (0) − ∆u1κ − κ∆u2κ uniformly bounded in H 1 (Ω);

∂u1κ
2. − gκt (0) uniformly bounded in H 1.5 (∂Ω), and
∂N

3. κ ku1κ kH 3 (Ω) + ku2κ kH 3 (Ω) is uniformly bounded.
An easy way to achieve the first two requirements is to let u3κ = fκt (0) −
∂u1κ
∆u1κ − κ∆u2κ in Ω and = gκt (0) on ∂Ω. We note that the choice of
∂N
u2κ could be made quite arbitrarily as long as
u2κ → u2 ≡ f (0) + ∆u0 in H 1 (Ω) and κku2κ kH 3 (Ω) ≤ Cku2 kH 1 (Ω) .
After all the analysis above, we state how u0κ , u1κ , u2κ , fκ and gκ
are defined in order to obtain κ-independent estimates. With {ζm }K
m=1
and {θm }K
m=1 denoting the partition of unity and local charts defined in
the proof of elliptic regularity, and Λ denoting the horizontal convolution

operator defined in Section 5.8, we define u0κ , fκ , and gκ by
u2κ = ηκ1/3 ∗ (E1 u2 ), fκ = ηκ1/3 ∗ (E1 f ),

K p
X p i −1
gκ = ζm Λκ1/3 ∗ ( ζm g) ◦ θm ◦ θm .
m=1
Then by the property of the convolution,

h i
κkfκ (0)kH 3 (Ω) ≤ C ku0 kH 3 (Ω) + kf (0)kH 1 (Ω) ,
kgκt (0)kH 1.5 (∂Ω) ≤ Cκ−1/3 kgt (0)kH 0.5 (∂Ω) ≤ Cκ−1/3 ku1 kH 2 (Ω) ,
and as κ → 0,
fκ → f in FΩ1 (T ), 1.5
gκ → g in G∂Ω (T ).
It remains to construct u0κ and u1κ . This can be done by the following
Lemma 7.12. Let f ∈ H k (Ω) and g ∈ H k−1.5 (∂Ω) for some k ≥ 1, and
∂f
= g on ∂Ω. Suppose that fn and gn are sequence of functions such that
∂N
fn → f in H k (Ω), gn → g in H k−1.5 (∂Ω), and
kfn kH k+` (Ω) ≤ K1 (n, `)kf kH k (Ω) , kgn kH k+`−0.5 (∂Ω) ≤ K2 (n, `)kgkH k−1.5 (∂Ω) .
Then the solution un to
un − λ n ∆un = fn in Ω, (7.55a)
∂un
= gn on ∂Ω, (7.55b)
∂N
belongs to H k+` (Ω), and

p
kun kH k+` (Ω) + λ n kun kH k+`+1 (Ω)
q h p i (7.56)
≤ C K21 + K22 kf kH k (Ω) + λ n kgkH k−1.5 (∂Ω) .
Suppose further that λ n > 0 and λ n → 0 as n → 0. Then
1. If for some ` ≥ 0, K1 and K2 are constants independent of n, then un

converges weakly to f in H k (Ω).
2. un converges (strongly) to f in H k (Ω) as n → ∞, provided that
λ n kun kH k+2 (Ω) → 0 as n → ∞. (7.57)
Proof. By the elliptic regularity,

h i
kun k2H k+` (Ω) + λ n kun k2H k+`+1 (Ω) ≤ C kfn k2H k+` (Ω) + λ n kgn k2H k+`−0.5 (∂Ω)
h i
≤ C K21 kf k2H k (Ω) + λ n K22 kgk2H k−1.5 (∂Ω)
h i
≤ C K21 + K22 kf k2H k (Ω) + λ n kgk2H k−1.5 (∂Ω) .
This proves (7.56).

Now suppose that for some ` ≥ 0, K1 and K2 are constants independent
of n. Then (7.56) implies that un is uniformly bounded in H k+` (Ω). In
particular, un is uniformly bounded in H k (Ω). Therefore, there exists a
subsequence nj of n and u ∈ H k (Ω) such that
unj * u in H k (Ω).
√
However, (7.56) also implies that λ n un is uniformly bounded in H 1 (Ω).
Passing j to the limit in the variational formulation
(unj , ϕ)L2 (Ω) + λ nj (Dunj , Dϕ)L2 (Ω)

= (fnj , ϕ)L2 (Ω) + λ nj (gnj , ϕ)L2 (∂Ω) ∀ ϕ ∈ H 1 (Ω),
we find that that (u, ϕ)L2 (Ω) = (f, ϕ)L2 (Ω) for all ϕ ∈ H 1 (Ω). This shows
u = f . This procedure also shows that f is the only weak limit of the whole
sequence un ; thus un converges weakly to f in H k (Ω).
If λ n kun kH k+2 (Ω) → 0 as n → ∞, then un − fn → 0 in H k (Ω) as n → ∞
since un − fn = λ n ∆un .
With Lemma 7.12, we then define u0κ as the solution to
(1 − κ∆)u0κ = ηκ1/3 ∗ (E3 u0 ) in Ω,

∂u0κ
= gκ (0) on ∂Ω .
∂N
and define u1κ as the solution to
(1 − κ∆)u1κ = ηκ1/3 ∗ (E2 u1 ) in Ω,

∂u1κ
= gκt (0) on ∂Ω .
∂N
Then by Lemma 7.12 (with k = 2 and ` = 0),

√ h √ i
ku1κ kH 2 (Ω) + κku1κ kH 3 (Ω) ≤ C kηκ1/3 ∗ (E2 u1 )kH 2 (Ω) + κkgκt (0)kH 0.5 (∂Ω)
√
≤ C(1 + κ)ku1 kH 2 (Ω) ,
Moreover, as κ → 0,
√
κkηκ1/3 ∗ (E3 u0 )kH 4 (Ω) ≤ Cκ1/6 ku0 kH 3 (Ω) → 0,
√
κkηκ1/3 ∗ (E2 u1 )kH 3 (Ω) ≤ Cκ1/6 ku1 kH 2 (Ω) → 0;
hence u0κ → u0 in H 3 (Ω) and u1κ → u1 in H 2 (Ω). Therefore, the study of
wt − κ∆w = fκ + ∆uκ in Ω × (0, T ) ,

∂w ∂uκ ∂u1κ ∂u2κ
κ + = gκ + κ + κt on ∂Ω × (0, T ) ,
∂N ∂N ∂N ∂N
j−1
∂t w = ujκ for j = 1, 2, 3 on Ω × {t = 0} ,
Z t
κ
u (t) = u0κ + w(s) ds,
0
implies that
h i
max kuκ (t)kH 3 (Ω)+ kuκt (t)kH 2 (Ω)+ kuκtt (t)kH 1 (Ω)+ kuκttt (t)kL2 (Ω)
t∈[0,T ]
h i
≤ C ku0 kH 3 (Ω) + (1 + κ)ku1 kH 2 (Ω) + kf kW 1 (T ) + kgkG 1.5 (T ) ;
Ω ∂Ω
thus the weak limit u of uκsatisfies

h i
kukW 3 (T ;Ω) ≤ C ku0 kH 3 (Ω) + ku1 kH 2 (Ω) + kf kW 1 (T ) + kgkG 1.5 (T ) . (7.58)
Ω ∂Ω
Remark 7.13. As discussed, we can always conclude (7.52) no matter

whether the condition
∂u0κ
= gκ (0) on ∂Ω (7.59)
∂N
is satisfied or not. Nevertheless, we will still regularize u0 in the fashion
that (7.59) is satisfied for consistency.
Remark 7.14. Without regularizing g, by the second-order compatibility

∂u
1κ
condition the term − gt (0) in estimate (7.53) is the same as

∂N H 1.5 (∂Ω)
∂(u − u )
1κ 1
. Since u1 ∈ H 2 (Ω) but u1 6∈ H 3 (Ω), it is impossible

∂N
1.5
H (∂Ω)
that on one hand κku1κ kH 3 (Ω) is uniformly bounded (which requires that
∂(u − u )
1κ 1
u1κ ∈ H 3 (Ω)), while on the other hand is bounded

∂N
1.5
H (∂Ω)
(which implies that u1κ at best belongs to H 0.5 (∂Ω)).
The case that m > 2
Suppose that u0 ∈ H m+1 , u1 ∈ H m , f ∈ FΩm−1 (T ) and g ∈ G∂Ω

m−1.5
(T )
satisfy the first m-th order compatibility conditions
∂u`
= (∂t` g)(0) on ∂Ω, (7.60)
∂N
where u`+2 = (∂t` f )(0)+∆u` for ` = 0, 1, · · · , m−1. Then for 0 ≤ ` ≤ m+1,
h i
ku` kH m−` (Ω) ≤ C ku0 kH m (Ω) + ku1 kH m−1 (Ω) + kf kW m−1 (T ) . (7.61)
Ω
We regularize the forcing functions f and g using the mollifier as before:

let fκ and gκ be defined by
fκ = η ∗ (Ef ), (7.62a)
K
X p p i −1
gκ = ζm Λ ∗ ( ζm g) ◦ θm ◦ θm (7.62b)
m=1
for some = (κ) < 1 to be determined later. Here we assume (κ) = κγ

for some γ > 0, and the goal is to find γ > 0 small enough so that we can
obtain κ-independent estimates for uκ . In general, as long as (κ) converges
slowly enough to 0 as κ → 0, then κ-independent estimates can be obtained.
As in the previous case, let u0κ be the solution to
u0κ − κ∆u0κ = η ∗ (Em+1 u0 ) in Ω,

∂u0κ
= gκ (0) on ∂Ω.
∂N
Then by elliptic regularity,
√
ku0κ kH m+`+1 (Ω) + κku0κ kH m+`+2 (Ω)
h √ i
≤ C kη ∗ (Em+1 u0 )kH m+`+1 (Ω) + κkgκ (0)kH m+`+0.5 (∂Ω)
h √ i
≤ C −` ku0 kH m+1 (Ω) + κ−`−1 kg(0)kH m−0.5 (∂Ω)
≤ C−` ku0 kH m+1 (Ω) (7.63)
as long as γ ≤ 1/2 (and from now on we assume that γ ≤ 1/2). Similarly,

let u1κ be the solution to
u1κ − κ∆u1κ = η ∗ (Em u1 ) in Ω,

∂u1κ
= gκt (0) on ∂Ω,
∂N
then
√
ku1κ kH m+` (Ω) + κku1κ kH m+`+1 (Ω) ≤ C−` ku1 kH m (Ω) . (7.64)
For 2 ≤ ` ≤ m − 1, define u`κ as the solution of
u`κ − κ∆u`κ = (∂t`−2 fκ )(0) + ∆u(`−2)κ in Ω, (7.65a)

∂u`κ
= (∂t` gκ )(0) on ∂Ω. (7.65b)
∂N
The elliptic regularity implies that for all k ≥ 0,
√
ku`κ kH m−`+k+1 (Ω) + κku`κ kH m−`+k+2 (Ω)
h √ i
≤ C k(∂t`−2 fκ )(0) + ∆u(`−2)κ kH m−`+k+1 (Ω) + κk∂t` gκ (0)kH m−`+k+0.5 (∂Ω) .
For ` = 2 and ` = 3, by (7.63) and (7.64) we find that

√ h i
ku2κ kH m+k−1 (Ω) + ku3κ kH m+k−2 (Ω) + κ ku2κ kH m+k (Ω) + ku3κ kH m+k−1 (Ω)
h 3
X
≤ C−k ku0 kH m+1 (Ω) + ku1 kH m+1 (Ω) + k(∂t`−2 f )(0)kH m−`−1 (Ω)
`=2
i
+ k∂t` g(0)kH m−`−0.5 (∂Ω)
h i
≤ C−k ku0 kH m+1 (Ω) + ku1 kH m+1 (Ω) + kf kW m−1 (T ) ,
Ω
which further implies that

√ h i
ku4κ kH m+k−3 (Ω) + ku5κ kH m−k−4 (Ω) + κ ku4κ kH m+k−2 (Ω) + ku5κ kH m+k−3 (Ω)
h 5
X
≤ C ku2κ kH m+k−1 (Ω) + ku3κ kH m+k−2 (Ω) + k(∂t`−2 fκ )(0)kH m+k−`−1 (Ω)
`=4
i
+ k∂t` gκ (0)kH m+k−`−0.5 (∂Ω)
h i
≤ C−k ku0 kH m+1 (Ω) + ku1 kH m+1 (Ω) + kf kW m−1 (T ) .
Ω
By induction we then conclude that for 0 ≤ ` ≤ m − 1 and k ≥ 0,

√
h i (7.66)
≤ C−k ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) .
Ω
In other words, we regularize u` in such a fashion that one derivative scales

like a multiplication by −1 , and the smooth version u`κ satisfies (7.65b).
Finally, for ` = m, m + 1, let
u`κ = (∂t`−2 fκ )(0) + ∆u(`−2)κ + κ∆u(`−1)κ . (7.67)
Then by (7.66), for ` = m, m + 1,

h i
ku`κ kH k (Ω) ≤ C k(∂t`−2 fκ )(0)kL2 (Ω)+ ku(`−2)κ kH k+2 (Ω)+ κku(`−1)κ kH k+2 (Ω)
h i
≤ C−k ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) .
Ω
In other words, (7.66) also holds for the case ` = m, m + 1.

Having u`κ , fκ and gκ defined, as discussed in the previous section, we
have the following
Definition 7.15 (The parabolic regularization of (7.14)). The m-th order

parabolic regularization of the hyperbolic initial-boundary value problem
utt − ∆u = f in Ω × (0, T ) ,
∂u
=g on ∂Ω × (0, T ) ,
∂N
u = u0 , ut = u1 on Ω × {t = 0} ,
is defined as
wt − κ∆w− ∆u = fκ + δfκ in Ω × (0, T ) , (7.68a)

∂w ∂u
κ + = gκ + δgκ on ∂Ω × (0, T ) , (7.68b)
∂N ∂N
w = u1κ , u = u0κ on Ω × {t = 0} , (7.68c)
where
m−3
X t`
δfκ = u(`+2)κ − (∂t` fκ )(0) − ∆u`κ − κ∆u(`+1)κ , (7.69a)
`!
`=0
m−1
X t` ∂u(`+1)κ
δgκ = κ . (7.69b)
`! ∂N
`=0
Estimates
In the following, we always assume that 0 < κ 1 (so that 1 as well).

By (7.65a) and (7.66),
m−3
Xh i
k∂tk (δfκ )kH m−k−1 (Ω) ≤ κ k∆u(`+2)κ kH m−k−1 (Ω) + k∆u(`+1)κ kH m−k−1 (Ω)
`=k
h i
k−m+1
≤ Cκ ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) . (7.70)
Ω
Moreover, (7.66) implies that for 0 ≤ k ≤ m,
k∂tk (δgκ )kH m−k−0.5 (∂Ω)

m−1 m−1
X ∂u(`+1)κ
X
≤ Cκ ≤ Cκ ku(`+1)κ kH m−k+1 (Ω)
∂N
m−k−0.5
H (∂Ω)
`=k `=k
m−1
X h i
≤ Cκ k−`−1 ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T )
Ω
`=k
h i
≤ Cκk−m ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) . (7.71)
Ω
Estimates (7.70) and (7.71) suggest that if we let γ = (m + 1)−1 (that is,
(κ) = κ1/(m+1) ), then we have
k∂tk (δfκ )kH m−k−1 (Ω) + k∂tk (δgκ )kH m−k−0.5 (∂Ω)
h i (7.72)
≤ C ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T )
Ω
which converges to 0 as κ → 0. Moreover, this choice of also validate

condition (7.57) in Lemma 7.12; thus u`κ → u` in H m−`+1 (Ω) as κ → 0 for
` = 0, 1, · · · , m − 1. On the other hand, = κ1/(m+1) implies that
h i
κ k∆u(m−1)κ kH 1 (Ω) + k∆umκ kL2 (Ω)
h i
≤ Cκ−1 ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) → 0 as κ → 0;
Ω
thus umκ → um in H 1 (Ω) and u(m+1)κ → um+1 in L2 (Ω) as well.

Giving the estimates for the forcing functions δfκ and δgκ , we are now
in the position of performing the estimates for the solution w to (7.68). For
1 ≤ k ≤ m, differentiating (7.68) in time k-times we obtain that
wkt − κ∆wk− ∆wk−1 = ∂tk fκ + ∂tk (δfκ ) in Ω × (0, T ) , (7.73a)

∂wk ∂wk−1
κ + = ∂tk gκ + ∂tk (δgκ ) on ∂Ω × (0, T ) , (7.73b)
∂N ∂N
wk = u(k+1)κ , wk−1 = ukκ on Ω × {t = 0}. (7.73c)
1. The case k = m:
Since ∂tm (δfκ ) = 0 and ∂tm (δgκ ) = 0, testing (7.73a) against wm , similar
to (7.31) we find that wm satisfies

h i Z T
2 2
sup kwm (t)kL2 (Ω) + kDwm−1 (t)kL2 (Ω) + κ kDwm (s)k2L2 (Ω) ds
t∈[0,T ] 0
h
≤ C ku(m+1)κ k2L2 (Ω) + kumκ k2H 1 (Ω) + k∂tm fκ k2L2 (0,T ;L2 (Ω))
+ k∂tm+1 fκ k2L2 (0,T ;L2 (Ω)) + k∂tm+1 gκ k2L2 (0,T ;H −0.5 (∂Ω))
i
+ k∂tm+2 gκ k2L2 (0,T ;H −0.5 (∂Ω))
h i
≤ C ku0 k2H m (Ω) + ku1 k2H m−1 (Ω) + kf k2W m−1 (T ) + kgk2W m−0.5 (T ) . (7.74)
Ω ∂Ω
By the fundamental theorem of Calculus (7.74) also implies that
max kwm−k (t)kL2 (Ω)

t∈[0,T ]
h i
≤ C ku0 kH m (Ω) + ku1 kH m−1 (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T )
Ω ∂Ω
for all k = 0, 1, · · · , m.
2. The case 1 ≤ k ≤ m − 1:
Since wkt = wk+1 , by Corollary 7.10 we find that that
h
max kwk−1 (t)kH m−k+1 (Ω) ≤ kukκ kH m−k+1 (Ω) + C max k∂tk fκ (t)kH m−k−1 (Ω)
t∈[0,T ] t∈[0,T ]
+ k∂tk (δfκ )(t)kH m−k−1 (Ω)+ k∂tk gκ (t)kH m−k−0.5 (∂Ω)+ k∂tk (δgκ )(t)kH m−k−0.5 (∂Ω)
i
+ kwk−1 (t)kL2 (Ω)
h i
≤ C ku0 kH m (Ω)+ ku1 kH m−1 (Ω)+ kf kW m−1 (T )+ kgkW m−0.5 (T ) . (7.75)
Ω ∂Ω
3. The case k = 0 or (7.68):

In this case, w0 = uκt , and we conclude from Corollary 7.10 again that
h
max kuκ (t)kH m+1 (Ω) ≤ ku0κ kH m+1 (Ω) + C max kfκ (t) + δfκ (t)kH m−1 (Ω)
t∈[0,T ] t∈[0,T ]
i
κ
+ kgκ (t) + δgκ (t)kH m−0.5 (∂Ω) + ku (t)kL2 (Ω)
h i
≤ C ku0 kH m (Ω) + ku1 kH m−1 (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T ) . (7.76)
Ω ∂Ω
Combining (7.74), (7.75), and (7.76), we obtain that
kuκ (t)kW m+1 (T ;Ω)

h i
≤ C ku0 kH m (Ω)+ ku1 kH m−1 (Ω)+ kf kW m−1 (T )+ kgkW m−0.5 (T ) .
Ω ∂Ω
7.3. THE PARABOLIC REGULARIZATION OF GENERAL HYPERBOLIC EQUATIONS303
Finally, since u`κ → u` , fκ → f , gκ → g, and δfκ → 0, δgκ → 0, we know

that the weak limit u of uκ is the unique solution to (7.14), and satisfies
kukW m+1 (T ;Ω)

h i
≤ C ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T ) .
Ω ∂Ω
(7.77)
7.3 The parabolic regularization of general hyper-

bolic equations
In the previous section, the parabolic regularization of the wave equation

with Neumann boundary condition is defined. The key is to introduce an ar-
tificial diffusion/viscosity so that the hyperbolic system becomes a parabolic
system. Moreover, if the initial data u0 , u1 or the forcing functions f, g pos-
sess limited regularity, we also need to regularize the data in such a way
that the the compatibility conditions for the new parabolic system (using
the regularized initial data as new initial data and the regularized forcing as
new forcing) are still satisfied. We remark that the compatibility conditions
for the parabolic regularization induces the compatibility conditions for the
hyperbolic equation: since (7.27b) holds for all κ > 0 at t = 0, by (7.28b)
we have
∂u` ∂`g
= ` (0) on ∂Ω, ` = 0, 1, · · · , m − 1.
∂N ∂t
If the coefficients aij is not constant, similar to Definition 6.26 we have
the following
Definition 7.16 (Compatibility conditions). For k ∈ N ∪ {0}, the (k +

1)th-order compatibility condition for the hyperbolic initial-boundary value
problem (7.1a,b,c,d) (resp. (7.1a,b,c,e)) is given by τ uk = 0 on ∂Ω
k−1

ij
resp. a (0) Ni = k (0) − Ni k−`
(0) on ∂Ω ,
∂xj ∂t ` ∂t ∂xj
`=0
where u` , ` = 2, · · · , k, is defined by
∂ `
u`+2 = (f + Lu) ` = 0, 1, · · · , k − 2.
∂t` t=0
Definition 7.16 induces the following
Definition 7.17 (Parabolic regularization). If the coefficients aij , bi , c of

the elliptic operator L, the initial data u0 , u1 , and the forcing functions f ,
g are smooth, then the m-th order parabolic regularization of the hyperbolic
equation
utt + Lu = f in Ω × (0, T ) , (7.78a)

∂u
aij Ni = g on ∂Ω × (0, T ) , (7.78b)
∂xj
u = u0 , ut = u1 on Ω × {t = 0} , (7.78c)
is defined as
Z t
u = u0 + w(s)ds in Ω × (0, T ) , (7.79a)
0
∂ ij ∂w
wt − κ a = f + Lu in Ω × (0, T ) , (7.79b)
∂xi ∂xj
∂w ∂u
κ aij Ni = g − aij Ni + δg on ∂Ω × (0, T ) , (7.79c)
∂xj ∂xj
w = u1 on Ω × {t = 0} , (7.79d)
where
m−1 `
t` ` ∂uk+1 ∂ `−k
XX
δg = κNi aij .
`! k ∂xj ∂t`−k t=0
`=0 k=0
When the coefficient aij , bi , c, initial data u0 , u1 and the forcing functions
f, g posses limited regularity, as discussed in Section 7.2.3, a regularization
of the data is required. We regularize the forcing functions f and g as in
(7.62), and regularize the data u` , for 0 ≤ ` ≤ m − 1, by elliptic equations
so that u`κ is the solution to
∂ ij ∂u`κ
u`κ − κ a (0) = F`κ in Ω, (7.80a)
∂xi κ ∂xj
∂u`κ
aij
κ (0) Ni = G`κ on ∂Ω, (7.80b)
∂xj
where aij ij
κ is a smooth version of a , G0κ = gκ (0), and
F`κ = η ∗ (Em+1−` u` ) if ` = 0, 1,
F`κ = ∂t`−2 t=0 (fκ + Lκ uκ )

if 2 ≤ ` ≤ m − 1,
`−1 `−r ij
X ` ∂ aκ ∂urκ
G`κ = (∂t` gκ )(0) − Ni (0) if 1 ≤ ` ≤ m − 1,
r ∂t`−r ∂xj
r=0
here with biκ , cκ denoting smooth versions of bi and c, respectively, the elliptic
operator Lκ is defined by
∂ ij ∂v ∂v
Lκ v ≡ − a + biκ + cκ v.
∂xi κ ∂xj ∂xi
We note that aij i

κ , bκ and cκ can be obtained by mollification (using parameter
= (κ)) as well. The reason for regularizing the coefficients is to make sure
that u`κ is smooth; for otherwise if aij ∈ C m (Ω × [0, T ]), no matter how
smooth the forcing functions are, u`κ at best belongs to H m+1 (Ω).
Similar to (7.67), for ` = m, m + 1 we define
u`κ = (∂t`−2 fκ )(0) − ∂t`−2 Lκ uκ − κ(aij κ

κ ut ,j ),i . (7.81a)
Definition 7.18 (Parabolic regularization). Let u`κ be defined by (7.80) and

(7.81). The m-th order parabolic regularization of the hyperbolic problem
utt + Lu = f in Ω × (0, T ) ,
∂u
aij Ni = g on ∂Ω × (0, T ) ,
∂xj
u = u0 , ut = u1 on Ω × {t = 0} ,
is defined as
Z t
κ
u = u0κ + w(s)ds in Ω × (0, T ) , (7.82a)
0
∂ ij ∂w
wt − κ a = fκ − Lκ uκ + δfκ in Ω × (0, T ) , (7.82b)
∂xi κ ∂xj
∂w ∂uκ
κ aij
κ Ni = gκ − aij
κ Ni + δgκ on ∂Ω × (0, T ) , (7.82c)
∂xj ∂xj
w = u1κ on Ω × {t = 0}. (7.82d)
where
m−2
X t`
u(`+2)κ − ∂t` t=0 fκ − Lκ uκ + κ(aij κ

δfκ = u ,
κ t j i ), ,
`!
`=0
m−1 `
t` ` ∂u(r+1)κ ∂ `−r
XX
δgκ = κNi aij .
`! r ∂xj ∂t`−r t=0 κ
`=0 r=0
7.3.1 Estimates for the regularized quantities
By the property of convolutions, we assume that
−k ij
1. kaij
κ kC m+k (Ω×[0,T ]) ≤ C ka kC m (Ω×[0,T ]) ;
2. kbiκ kC m+k (Ω×[0,T ]) ≤ C−k kbi kC m (Ω×[0,T ]) ;
3. kcκ kC m+k (Ω×[0,T ]) ≤ C−k kckC m (Ω×[0,T ]) ;
4. kfκ kW m+k−1 (T ) ≤ C−k kf kW m−1 (T ) ;

Ω Ω
5. kgκ kG m+k−0.5 (T ) ≤ C−k kgkW m−0.5 (T ) .

∂Ω ∂Ω
To simplify the notation, throughout the rest of the section we let
N (u0 , u1 , f, g) = ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T ) .

Ω ∂Ω
Estimates for u`κ
The estimates for u`κ is much more complicated than the case that aij = δij ,
so we derive the estimates in great detail. As before, we always assume that
0 < κ 1 and = (κ) = κγ for some γ < 0.5 to be determined later.
Lemma 7.19. Suppose that ∂Ω is smooth, and aij i

κ , bκ , cκ , fκ and gκ are
defined by mollification, and satisfy the regularity assumptions 1-5. Let u`κ
be defined by (7.80) and (7.81). Then for 0 ≤ ` ≤ m + 1 and 0 ≤ k ≤ Km,
√
ku`κ kH m−`+k+1 (Ω) + κku`κ kH m−`+k+2 (Ω) ≤ C−k ×
h i
× ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T ) ,
Ω ∂Ω
(7.83)
here the constant C = C kakC m (Ω×[0,T ]) , λ 0 , K).
Proof. Let 0 ≤ ` ≤ m be fixed. Suppose that we can show that for k ≥ 0,
kF`κ kH m−`+k+1 (Ω) ≤ C−k N (u0 , u1 , f, g) (7.84)
and
√
κkG`κ kH m−`+k+0.5 (∂Ω) ≤ C−k N (u0 , u1 , f, g), (7.85)
then by elliptic regularity, for k ≥ 0,

√
m−`+k+1
√ X
≤C κ kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω)
s=1
h √ i
+ C kF`κ kH m−`+k+1 (Ω) + κkG`κ kH m−`+k+0.5 (∂Ω)
√ m−`+k+1
X
≤C κ kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω) (7.86)
s=1
+ C−k N (u0 , u1 , f, g)
for some C = C kaij

κ (0)kC m (Ω) , λ 0 , K .
1. If 0 ≤ k ≤ ` − 1, then kaκ (0)kC s (Ω) ≤ CkakC m (Ω×[0,T ]) . Therefore,

m−`+k+1
X
kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω) ≤ Cmku`κ kH m−`+k+1 (Ω)
s=1

for some constant C = C kakC m (Ω×[0,T ]) . By κ 1, (7.84) and (7.85)
imply that
√
κku`κ kH m−`+k+2 (Ω) ≤ C−k N (u0 , u1 , f, g),
ku`κ kH m−`+k+1 (Ω) +
(7.87)

where C = C kakC m (Ω×[0,T ]) , λ 0 .
2. If ` ≤ k ≤ ` − 1 + m, then
m−`+k+1
X
kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω)
s=1
m
X m−`+k+1
X
= + kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω)
s=1 s=m+1
m−`+k+1
X
≤ Cmku`κ kH m−`+k+1 (Ω) + C m−s ku`κ kH m−`+k−s+2 (Ω) .
s=m+1
By κ 1, we find that
√
√ m−`+k+1
X
≤C κ m−s ku`κ kH m−`+k−s+2 (Ω) + C−k N (u0 , u1 , f, g).
s=m+1
If m + 1 ≤ s and k ≤ ` − 1 + m, then
k − s ≤ ` − 2 ≤ ` − 1;
thus by (7.87),
√
κku`κ kH m−`+k−s+2 (Ω) ≤ Cs−k N (u0 , u1 , f, g).
As a consequence,
√
h m−`+k+1
X i
≤C m−s s−k + −k N (u0 , u1 , f, g) ≤ C−k N (u0 , u1 , f, g).
s=m+1
Therefore, (7.83) holds if 0 ≤ k ≤ ` + m.
3. Suppose that (7.83) holds for the case ` + (p − 1)m ≤ k ≤ ` − 1 + pm.

Then if ` + pm ≤ k ≤ ` − 1 + (p + 1)m (and k ≤ Km),
m−`+k+1
X
kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω)
s=1
m
X m−`+k+1
X
= + kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω) .
s=1 s=m+1
If s ≥ m + 1, then k − s ≤ ` − 1 + pm; thus
m−`+k+1
X √
kaκ (0)kC s (Ω) κku`κ kH m−`+k−s+2 (Ω)
s=m+1
m−`+k+1
X
≤C m−s −(k−s) N (u0 , u1 , f, g) ≤ C−k N (u0 , u1 , f, g)
s=m+1
for some constant C = C kakC m (Ω×[0,T ]) , λ 0 , K). Therefore, (7.86)

with κ 1 implies that
√
√ m−`+k+1
X
≤C κ kaκ (0)kC s (Ω) ku`κ kH m−`+k−s+2 (Ω) + C−k N (u0 , u1 , f, g)
s=m+1
−k
≤ C N (u0 , u1 , f, g);
thus (7.83) holds if ` + pm < k ≤ ` + (p + 1)m. By induction, (7.83)

holds for 0 ≤ k ≤ Km if (7.84) and (7.85) hold.
Now we prove (7.84) and (7.85) for all ` ≥ 0.
1. If ` = 0, (7.84) and (7.85) trivially hold because of the regularity

assumptions on fκ and gκ . Therefore, (7.83) holds for the case ` = 0.
In particular,
ku0κ kH k (Ω) ≤ C max 1, m+1−k N (u0 , u1 , f, g).

(7.88)
2. If ` = 1, (7.84) holds trivially. For (7.85), we note that if h ∈ C 1 (∂Ω)

and v ∈ H 0.5 (∂Ω), then
khvkH 0.5 (∂Ω) ≤ CkhkC 1 (∂Ω) kvkH 0.5 (∂Ω) .
By kaκ kC k (Ω×[0,T ]) ≤ C max 1, m−k kakC m (Ω×[0,T ]) and (7.88),

∂aij ∂u0κ
κ
Ni (0)

∂t ∂xj H m+k−0.5 (∂Ω)

m+k−1
X
≤C kaκt (0)kC s+1 (∂Ω) kDu0κ kH m+k−s−0.5 (∂Ω)
s=0
m+k−1
X
≤C kaκ kC s+2 (Ω×[0,T ]) ku0κ kH m+k−s+1 (Ω)
s=0
m+k−1
X
max 1, m−s−2 max 1, s−k N (u0 , u1 , f, g).

≤C
s=0

for some constant C = C kakC m (Ω×[0,T ]) . If 0 ≤ s ≤ m + k − 1,
m − s − 2 ≥ −k − 1 , s − k ≥ −k − 1 and m − k − 2 ≥ −k − 1 ;
thus by max{a, b} max{c, d} ≤ max{ac, bc, ad, bd} if a, b, c, d ≥ 0,
max 1, m−s−2 max 1, s−k ≤ max 1, m−s−2 , s−k , m−k−2 ≤−k−1 .

√
Therefore, by κ−1 ≤ 1,
ij
√ ∂aκ ∂u0κ
κ Ni (0) ≤ C−k N (u0 , u1 , f, g),

∂t ∂xj H m+k−0.5 (∂Ω)

which implies that (7.85) holds for ` = 1. This also implies that (7.83)
holds for ` = 1, and
ku1κ kH k (Ω) ≤ C max 1, m−k N (u0 , u1 , f, g).

(7.89)
3. Suppose that (7.84) and (7.85) hold for all 0 ≤ ` ≤ k − 1 for some
p ≥ 2. Then similar to (7.88) and (7.89), for all ` ≤ (p − 1) and k ≥ 0,
ku`κ kH k (Ω) ≤ C max{1, m−k−`+1 N (u0 , u1 , f, g).

(7.90)
To estimate Fpκ in H m−p+k+1 (Ω), we first note that by Liebnitz rule,
k∂tp−2 t=0 (Lκ uκ )kH m−p+k+1 (Ω)

p−2 m−p+k+1
X X h
≤C kaκ kC p+s−r−2 (Ω×[0,T ]) kurκ kH m−p−s+k+3 (Ω)
r=0 s=0
+ kbκ kC p+s−r−2 (Ω×[0,T ]) kurκ kH m−p−s+k+2 (Ω) (7.91)
i
+ kcκ kC p+s−r−2 (Ω×[0,T ]) kurκ kH m−p−s+k+1 (Ω) .
By (7.90),
p−2 m−p+k+1
X X
kaκ kC p+s−r−2 (Ω×[0,T ]) kurκ kH m−p−s+k+3 (Ω)
r=0 s=0
p−2 m−p+k+1
X X m−p−s+r+2 p+s−k−r−2 m−k
≤ max 1, , , N (u0 , u1 , f, g).
r=0 s=0
However, if 0 ≤ s ≤ m − p + k + 1, then
m − p − s + r + 2 > −k and p + s − k − r − 2 > −k ;
thus
max 1, m−p−s+r+2 , p+s−k−r−3 , m−k ≤ −k .

The other two terms in (7.91) can be estimated in the same fashion.
Therefore,
k∂tp−2 t=0 (Lκ uκ )kH m−p+k+1 (Ω) ≤ C−k N (u0 , u1 , f, g),

(7.92)
which implies that (7.84) holds for the case ` = k.
Similarly,
`−1 `−r ij
X ∂ aκ ∂urκ
Ni (0)

∂t`−r ∂xj H m−`+k+0.5 (∂Ω)

r=0
`−1 m−`+k
X X
≤C k(∂t`−r aij
κ )(0)kC s+1 (Ω) kDurκ kH m−`+k−s+0.5 (∂Ω)
r=0 s=0
`−1 m−`+k
X X
≤C kaκ kC `−r+s+1 (Ω×[0,T ]) kurκ kH m−`+k−s+2 (Ω)
r=0 s=0
`−1 m−`+k
X X
max 1, m−`+r−s−1 , `−k+s−r−1 , m−k−2 N (u0 , u1 , f, g).

≤C
r=0 s=0
If 0 ≤ s ≤ m − ` + k and 0 ≤ r ≤ ` − 1, then
m − ` + r − s − 1 ≥ −k − 1 and ` − k + s − r − 1 ≥ −k − 1 ;
thus
max 1, m−`+r−s−1 , `−k+s−r−1 , m−k−2 ≤ −k−1 .

As a consequence,
`−1 `−r ij
X ∂ aκ ∂urκ
Ni (0) ≤ C−k−1 N (u0 , u1 , f, g)

∂t`−r ∂xj H m−`+k+0.5 (∂Ω)

r=0
which implies that
√
κkG`κ kH m−`+k+0.5 (∂Ω) ≤ C−k N (u0 , f, g).
The estimates for the case ` = m and ` = m + 1 are similar, and are
left as an exercise.

Estimates for δfκ and δgκ
Lemma 7.19 establishes estimates corresponding to (7.66) in the case of wave

equations. In order to show that the solution to the parabolic regularization
has κ-independent estimates, estimates corresponding to (7.70) and (7.71)
have to be established as well. Similar to (7.70) and (7.71), we have the
following
Lemma 7.20. Let δfκ and δgκ be given in Definition 7.18. Then
k∂tk (δfκ )kH m−k−1 (Ω) + k∂tk (δgκ )kH m−k−0.5 (∂Ω) ≤ Cκk−m ×
h i
× ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T ) .
Ω ∂Ω
(7.93)
Proof. We recall that
m−3
X t`
u(`+2)κ − ∂t` t=0 fκ − Lκ uκ + κ(aij κ

δfκ = κ ut ,j ),i ,
`!
`=0
m−1 `
t` ` ∂u(r+1)κ `−k ij
XX
δgκ = κNi (∂t aκ )(0) .
`! r ∂xj
`=0 r=0
We note that for ` ≤ m − 2,
`
∂t (aij uκ , ), ≤ ∂t` t=0 (aij κ

t=0 κ t j i H m−k+1 (Ω) κ ut ,j ) H m−k+2 (Ω)

` m−k+2
X X
≤ k(∂t`−r aκ )(0)kC s (Ω) ku(r+1)κ kH m−k−s+3 (Ω)
r=0 s=0
X m−k+2
` X
max 1, m−`+r−s , k+s−r−3 , m−`+k−3 N (u0 , u1 , f, g)

≤C
r=0 s=0
≤ Ck−m−1 N (u0 , u1 , f, g). (7.94)
Therefore, by (7.80a),
m−3
X
k∂tk (δfκ )kH m−k−1 (Ω) u(`+2)κ − ∂t` (fκ − Lκ uκ m−k+1

≤C t=0 H (Ω)
`=k
m−3
X `
∂t (aij uκ , ),

+ Cκ t=0 κ t j i H m−k+1 (Ω)
;
`=k
thus (7.94) implies that
k∂tk (δfκ )kH m−k−1 (Ω)

m−3
X ij
(aκ (0)u(`+2)κ ,j ),i m−k−1 + Cκk−m−1 N (u0 , u1 , f, g)

≤ Cκ H (Ω)
`=k
k−m
≤ Cκ N (u0 , u1 , f, g).
Similarly,
k∂tk (δgκ )kH m−k−0.5 (∂Ω)

m−1
XX ` m−k−1
X
≤ Cκ k(∂t`−r aij
κ )(0)kC s+1 (∂Ω) kDu(r+1)κ kH m−k−s−0.5 (∂Ω)
`=k r=0 s=0
m−1
XX ` m−k−1
X
≤ Cκ kaij
κ kC `−r+s+1 (Ω×[0,T ]) ku(r+1)κ kH m−k−s+1 (Ω)
`=k r=0 s=0
m−1
XX ` m−k−1
X
max 1, m−`+r−s−1 max 1, k+s−r−1 N (u0 , u1 , f, g)

≤ Cκ
`=k r=0 s=0
m−1
XX ` m−k−1
X
max 1, m−`+r−s−1 , k+s−r−1 , m−`−2 N (u0 , u1 , f, g)

≤ Cκ
`=k r=0 s=0
≤ Cκk−m N (u0 , u1 , f, g),
where the last inequality follows from that m − ` + r − s − 1 ≥ k − m,

k + s − r − 1 ≥ k − m, and m − ` − 2 ≥ k − m if k ≤ ` ≤ m − 1, 0 ≤ r ≤ `,
and 0 ≤ s ≤ m − k − 1.
1
Finally, choose γ = . Then Lemma 7.20 implies that ∂tk (δfκ ) → 0
m+1
in H m−k−1 (Ω) and ∂tk (δgκ ) → 0 in H m−k−0.5 (∂Ω). Moreover, with this
choice of γ, we have the following
Lemma 7.21. Let = κ1/(m+1) . Then u`κ → u` in H m−`+1 (Ω) for all
0 ≤ ` ≤ m + 1.
Proof. First of all, by (7.80a),
ku`κ − F`κ kH m−`+1 (Ω) = κ (aij

κ (0)u`κ ,j ),i H m−`+1 (Ω)

m−`+2
X
≤ Cκ kaij
κ (0)kC s (Ω) ku`κ kH m−`−s+2 (Ω)
s=0
m−`+2
X
max 1, m−s , s−2 , m−2 N (u0 , u1 , f, g)

≤ Cκ
s=0
1−m
≤ Cκ N (u0 , u1 , f, g);
thus u`κ − F`κ → 0 in H m−`+1 (Ω). The conclusion then follows from the
fact that F`κ → u` in H m−`+1 (Ω) for all 0 ≤ ` ≤ m + 1 by the property of
convolution.
1
Remark 7.22. In Lemma 7.21 we could have chosen = − , in which
log κ
case for all m ∈ N,
u`κ → u` in H m−`+1 (Ω),

∂tk (δfκ ) → 0 in H m−k−1 (Ω),
∂tk (δgk ) → 0 in H m−k−0.5 (∂Ω),
as κ → 0.
7.3.2 Main theorem
Theorem 7.23. Assume that ∂Ω is of class C ∞ , ∂tt + L is uniformly hy-

perbolic, aij , aij i m ij
t , b , c ∈ C (Ω × [0, T ]), and [a ] is symmetric. Then for any
u0 ∈ H m+1 (Ω), u1 ∈ H m (Ω), f ∈ FΩm−1 (T ), and g ∈ G∂Ω
m−0.5
(T ), the unique
weak solution u to the hyperbolic initial-boundary value problem (7.1a,b,c,e)
in fact belongs to W m+1 (T ; Ω) provided that u0 satisfies the compatibility
condition up to m-th order; that is,
k−1

aij (0) Ni = k (0) − Ni k−`
(0) on ∂Ω
`=0
for k = 1, · · · , m. Moreover, u satisfies

kukW m+1 (T ;Ω)
h i
≤ C ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T )
Ω ∂Ω
(7.77)
7.4. AN ALTERNATIVE APPROACH - THE GALERKIN METHOD315

where C = C k(a, at , b, c)kC m (Ω×[0,T ])4 , λ 0 , T .
The proof of Theorem 7.23 is essentially the same as for the case of wave
equations, which involves studying the m-times time-differentiated version
of the problem, and building up regularity using Corollary 7.10. The detail
is left as an exercise.
7.4 An alternative approach - the Galerkin method

7.4.1 Existence of weak solutions
Similar to the parabolic case, we can apply the Galerkin method to construct
a weak solution to equation (7.1a,b,c,e). In the following, we assume that
k
dk` (t)e` (t) be the finite dimensional
P
g = 0 for simplicity. Let uk (x, t) =
`=1
approximation of u(x, t) satisfying

uktt (t), ϕ L2 (Ω)+ B uk (t), ϕ = f (t), ϕ L2 (Ω) ∀ ϕ ∈ span(e1 , · · · , ek ) .
(7.95)
Solving the system above is equivalent to solving an ODE system

k
M`m (t)dk` (t) = fˆm (t) ,
X
dkm00 (t) + (7.96a)
`=1
dk` (0) = (u0 , e` )L2 (Ω) , dk` 0 (0) = (u1 , e` )L2 (Ω) , (7.96b)
where the matrix M`m is defined as before. The ODE system above has a
unique solution dk (t) in (0, T ) as long as M ∈ L∞ (0, T ), and it follows that
uk exists on the time interval [0, T ]. The key here is to derive a uniform
bound for uk , ukt and uktt in appropriate spaces.
The estimates for uk and ukt
Unlike the parabolic case, in order to estimate uk , we plug in ϕ = ukt instead

of uk to obtain
1d ∂uk ∂ukt
Z Z
2 ij
kukt (t)kL2 (Ω) + a (x, t) (x, t) (x, t) dx= f (x, t)ukt (x, t) dx
2 dt Ω ∂xj ∂xi Ω
∂u
Z h i
k
− bi (x, t) (x, t)ukt (x, t) + c(x, t)uk (x, t)ukt (x, t) dx . (7.97)
Ω ∂xi
Therefore, (6.23) together with the Hölder and Young inequalities implies
that
1 dh ∂uk ∂uk
Z i
2
kukt (t)kL2 (Ω) + aij (x, t) (x, t) (x, t) dx
2 dt Ω ∂xj ∂xi
1
≤ kat (t)kL∞ (Ω) kDuk (t)k2L2 (Ω)
2
1 h i
+ kb(t)kL∞ (Ω) kDuk (t)k2L2 (Ω) + kukt (t)k2L2 (Ω) (7.98)
2
1 h i
+ kc(t)kL∞ (Ω) kuk (t)k2L2 (Ω) + kukt (t)k2L2 (Ω)
2
1h i
+ kf (t)k2L2 (Ω) + kukt (t)k2L2 (Ω) .
2
Z t
By the fundamental theorem of calculus, uk (t) = uk (0) + ukt (s) ds , so
0
hZ s i2
kuk (s)k2L2 (Ω) ≤ 2kuk (0)k2L2 (Ω) + 2 kukt (t0 )kL2 (Ω) dt0
Z0 s
≤ 2kuk (0)k2L2 (Ω) + 2s kukt (t0 )k2L2 (Ω) dt0 . (7.99)
0
Integrating (7.98) in time over the time interval (0, t), the use of (7.99) and
the uniform hyperbolicity suggest that
h
kukt (t)k2L2 (Ω) + λ 0 kDuk (t)k2L2 (Ω) ≤ C ku0 k2H 1 (Ω) + ku1 kL2 (Ω)
Z th i
2
+ kf kL2 (0,T ;L2 (Ω)) + kukt (s)k2L2 (Ω) + kDuk (s)k2L2 (Ω) ds ,
0
and we can further conclude from the Gronwall inequality that
kukt (t)k2L2 (Ω) + kDuk (t)k2L2 (Ω)

h i (7.100)
≤ C ku0 k2H 1 (Ω)+ ku1 k2L2 (Ω)+ kf k2L2 (0,T ;L2 (Ω))

for some constant C = C kakC 1 (Ω×[0,T ] , kbkL∞ (Ω×(0,T )) , kckL∞ (Ω×(0,T )) , λ 0 , T .
We note that if aij is time-independent, the dependence of C on kakC 1 (Ω×[0,T ])
can be replaced by kakL∞ (Ω) .
Using (7.99) again to obtain an upper bound for kuk (t)k2L2 (Ω) , (7.100)
then suggests that
kukt (t)k2L2 (Ω) + kuk (t)k2H 1 (Ω)

h i (7.101)
≤ C ku0 k2H 1 (Ω)+ ku1 k2L2 (Ω)+ kf k2L2 (0,T ;L2 (Ω)) .
§7.4 An alternative approach - the Galerkin method 317
The estimates for uktt
Similar to how ukt is estimated in the parabolic case, for ϕ ∈ H 1 (Ω), we

express ϕ as the sum of ϕ1 and ϕ2 such that ϕ1 ∈ span(e1 , · · · , ek ) and

(ϕ1 , ϕ2 )H 1 (Ω) = 0 to obtain uktt (t), ϕ H 1 (Ω) = (uktt (t), ϕ1 )L2 (Ω) and con-
clude that
kuktt (t)kH 1 (Ω)0 ≤ kf (t)kL2 (Ω) + C(t)kuk (t)kH 1 (Ω) , (7.102)
where C(t) = ka(t)kL∞ (Ω) + kb(t)kL∞ (Ω) + kc(t)kL∞ (Ω) . The combination of
(7.101) and (7.102) implies that
kuk kLp (0,T ;H 1 (Ω)) + kukt kLp (0,T ;L2 (Ω)) + kuktt kL2 (0,T ;H 1 (Ω)0 )
h i (7.103)
≤ C ku0 kH 1 (Ω) + ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω))
for some constant C = C kakC 1 (Ω×[0,T ]) , kbkC(Ω×[0,T ]) , kckC(Ω×[0,T ]) , λ 0 , T )

independent of p.
Weak convergence
In order to apply Banach-Alaoglu theorem, the underlying space has to be

reflexive. By (7.103), uktt is uniformly bounded in L2 (0, T ; H 1 (Ω)0 ), while
Duk and ukt are uniformly bounded in Lp (0, T ; L2 (Ω)) for all 1 < p < ∞.
Therefore, there exists a subsequence kj of k and u ∈ Lp (0, T ; H 1 (Ω)), such
that for all p ∈ N,
ukj * u in Lp (0, T ; H 1 (Ω)), (7.104a)

ukj t * ut in Lp (0, T ; L2 (Ω)), (7.104b)
ukj tt * utt in L2 (0, T ; H 1 (Ω)0 ), (7.104c)
as j → ∞, and u satisfies the alternative variational formulation

Z T h
i
utt (t), v(t) H 1 (Ω) + B u(t), v(t) dt
0 Z T (7.105)
∀ v ∈ L2 (0, T ; H 1 (Ω))

= f (t), v(t) L2 (Ω) dt
0
which is equivalent to (7.6) (the proof is the same as the parabolic case).
Moreover, by the property of lower semi-continuity of norms, for all p ∈ N
(7.103) implies that the weak limit u satisfies
kukLp (0,T ;H 1 (Ω)) + kut kLp (0,T ;L2 (Ω)) + kutt kL2 (0,T ;H 1 (Ω)0 )
h i
≤ C ku0 kH 1 (Ω) + ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω))
for some constant C independent of p. By that lim khkLp (0,T ) = khkL∞ (0,T ) ,
p→∞
we conclude that
kukL∞ (0,T ;H 1 (Ω)) + kut kL∞ (0,T ;L2 (Ω)) + kutt kL2 (0,T ;H 1 (Ω)0 )
h i
≤ C ku0 kH 1 (Ω) + ku1 kL2 (Ω) + kf kL2 (0,T ;L2 (Ω)) ;
(7.106)
thus u ∈ L∞ (0, T ; H 1 (Ω)) with time derivatives ut ∈ L∞ (0, T ; L2 (Ω)) and
utt ∈ L2 (0, T ; H 1 (Ω)0 ).
Finally, let ζ ∈ C 2 ([0, T ]) with the property that ζ(0) = ζ 0 (0) = 1 and
ζ(T ) = ζ 0 (T ) = 0. Using v(x, t) = ζ(t)ϕ(x) for some ϕ ∈ H 1 (Ω) as a test
function in (7.105), integrating by parts twice in time we find that
Z Th
i
u(t), ζ 00 (t)ϕ L2 (Ω))+ B u(t), ζ(t)ϕ dt

− ut (0), ϕ)L2 (Ω)+ u(0), ϕ)L2 (Ω)+
0
Z Th
i
− ut (t), ζ 0 (t)ϕ L2 (Ω)) + B u(t), ζ(t)ϕ dt

= − ut (0), ϕ)L2 (Ω) +
0
Z T
∀ ϕ ∈ H 1 (Ω).

= f (t), ζ(t)ϕ L2 (Ω) dt (7.107)
0
Multiplying (7.95) by ζ, integrating by parts in time twice, and then passing

k → ∞, we obtain that
Z Th
i
u(t), ζ 00 (t)ϕ L2 (Ω))+ B u(t), ζ(t)ϕ dt

− u1 , ϕ)L2 (Ω)+ u0 , ϕ)L2 (Ω)+
0
Z Th
i
− ut (t), ζ 0 (t)ϕ L2 (Ω)) + B u(t), ζ(t)ϕ dt

= − u1 , ϕ)L2 (Ω) +
0
Z T
∀ ϕ ∈ H 1 (Ω).

= f (t), ζ(t)ϕ L2 (Ω) dt (7.108)
0
Comparing (7.107) to (7.108), we conclude that u(0) = u0 and ut (0) =

u1 . As a consequence, we establish the existence of a weak limit u to the
hyperbolic initial-boundary value problem (7.1a,b,c,e).
The detail derivation of (7.107) and (7.108) is similar to (6.19) and (6.20)
in the parabolic case, and is left as an exercise.
§7.4.2 Regularity theory revisited 319
7.4.2 Regularity theory revisited
Suppose that u0 ∈ H 2 (Ω), u1 ∈ H 1 (Ω), f ∈ L2 (0, T ; L2 (Ω)) with ft ∈

L2 (0, T ; L2 (Ω)), and aij , bi , c ∈ C 1 (Ω × [0, T ]). As for the parabolic case,
in order to show that the solution u to (7.1a,b,c,e) possesses better spatial
regularity, we tend to apply the elliptic regularity to the elliptic equation
Lu = f − utt in Ω,
∂u
aij Nj = 0 on ∂Ω ;
∂xi
thus result in the study of the regularity of utt . However, unlike the case
that one can test the parabolic equation against ut to obtain an estimate for
kut kL2 (Ω) , for the hyperbolic problem we do not have the luxury to test the
hyperbolic equations against utt in order to obtain meaningful estimates.
A naive idea is to obtain the regularity of utt by studying of the time-
differentiated version of the problem: let w = ut , and obtain the regularity
of wt by testing the time-differentiated version of the problem, a hyperbolic
equation of w, against wt . However, this process will lead to the boundary
condition
∂w ∂u
aij Ni = −aij
t Ni on ∂Ω (7.109)
∂xj ∂xj
in which the right-hand side does not make sense when a is time-dependent
and u ∈ L2 (0, T ; H 1 (Ω)). To resolve this issue, we focus on the equation
K
X
uttt − aij ut ,j ,i− aij u,j ,i = F + aij hk ,j ,i in Ω×(0, T ) , (7.110a)

k=1
K
∂ut ∂u X ∂hk
aij Ni + aij Ni = G− aij Ni on ∂Ω×(0, T ) ,
∂xj ∂xj ∂xj
k=1
(7.110b)
u = u0 , ut = u1 , utt = u2 on Ω×{t = 0} ,
(7.110c)
where aij are assumed to satisfy aij ξj ξi ≥ λ 0 |ξ|2 , and aij , aij , aij 1
k ∈ C (Ω ×
[0, T ]) for k = 1, · · · , K, F ∈ L2 (0, T ; L2 (Ω)), G, Gt ∈ L2 (0, T ; H −0.5 (∂Ω)),
and hk ∈ W 2 (T ; Ω) for k = 1, · · · , K. For the moment we assume that the
initial data possess the regularity u0 ∈ H 1 (Ω), u1 ∈ H 1 (Ω) and u2 ∈ L2 (Ω).
We note that (7.110) is the first time-differentiated version of (7.1) if aij =

aij i
t , b = c = 0, F = ft , G = gt , and hk = 0 for all k, and all the higher
order time-differentiated version of (7.1) can be written in the form (7.110)
with different u, aij , aij
k and hk .
Definition 7.24. A function u ∈ L∞ (0, T ; H 1 (Ω)) with time-derivatives

ut ∈ L∞ (0, T ; H 1 (Ω)), utt ∈ L2 (0, T ; L2 (Ω)) and utt ∈ L2 (0, T ; H 1 (Ω)0 ) is
said to be a weak solution to (7.110) provided that

uttt (t), ϕ H 1 (Ω) + B1 ut (t), ϕ + B2 u(t), ϕ = F (t), ϕ L2 (Ω) (7.111)
K Z
aij
X
1

+ G(t), ϕ H 0.5 (∂Ω) − k hk ,j ϕ,i dx ∀ ϕ ∈ H (Ω), a.e. t ∈ (0, T ) ,
k=1 Ω
and
u = u0 , ut = u1 , utt = u2 on Ω × {t = 0} , (7.112)

where B1 ut (t), ϕ is defined by
∂ut ∂ϕ
Z
aij (x, t)

B1 ut (t), ϕ ≡ (x, t) (x) dx,
∂xj ∂xi
ZΩ
∂u ∂ϕ
aij (x, t)

B2 u(t), ϕ ≡ (x, t) (x) dx.
Ω ∂xj ∂xi
Equation (7.111) is called the variational formulation of (7.110).
Since the estimates for all K are the same, we assume that K = 1
and write h1 = h. Let {e` }∞ 2
`=1 be an orthonormal basis of L (Ω) which
k
is orthogonal in H 1 (Ω), and let uk (x, t) =
P k
dm (t)em (x) be the finite
m=1
dimensional approximation of the solution, where dkm satisfies the ODE
k h
X i
dkm000 (t) + M`m (t)dk` 0 (t) + M̃`m (t)dk` (t) = F̂m (t) + Ĝm (t) + Ĥm (t),
`=1
(7.113a)
dk` (0)=(u0 , e` )L2 (Ω) , dk` 0 (0)=(u1 , e` )L2 (Ω) , dk` 00 (0)=(u2 , e` )L2 (Ω) ,
(7.113b)

where M`m (t) = B1 (e` , em ), M̃`m (t) = B2 (e` , em ), F̂m (t) = F (t), em L2 (Ω) ,

Ĝm (t) = G(t), e` H 0.5 (∂Ω) , and
Z
Ĥm (t) = aij (t)h,j (t)ϕ,i dx.
Ω
Note that (7.113a) is the same as
(ukttt , ϕ)L2 (Ω) + B1 (ukt , ϕ) + B2 (uk , ϕ) = (F, ϕ)L2 (Ω)

Z
+ hG, ϕiH 0.5 (∂Ω) − aij h,j ϕ,i dx ∀ ϕ ∈ span(e1 , · · · , ek ).
Ω
(7.114)
By the fundamental theorem of ODE, there exists a unique dk ∈ C 3 ([0, Tk ])
for some 0 < Tk ≤ T if aij , aij ∈ C(Ω × [0, T ]).
Let ϕ = uktt in (7.114), we find that
1 dh i
kuktt (t)k2L2 (Ω) + B1 ukt (t), ukt (t) = F (t), uktt (t) L2 (Ω)

2 dt
1 ∂ukt ∂ukt
Z
aij

+ t (t) (t) (t)dx + G(t), uktt (t) H 0.5 (∂Ω) (7.115)
2 Ω ∂xj ∂xi
∂h ∂uktt
Z
aij

− dx − B2 uk (t), uktt (t) .
Ω ∂xj ∂xi
The key to close the estimate is to estimate the time integral of the last two
terms on the right-hand side. Nevertheless, integrating by parts in time, we
find that
Z t
B2 uk (s), uktt (s) ds
0
Z tZ
s=t ∂ h ij ∂u (x, s) ∂ukt (x, s)
i
= B2 uk (s), ukt (s) − a (x, s) k dxds
s=0 0 Ω ∂s ∂xj ∂xi
h i
≤ C kDuk (t)k2L2 (Ω) + kuk (0)k2H 1 (Ω) + kukt (0)k2H 1 (Ω) + kDukt (t)k2L2 (Ω)
Z th i
+C kDuk (s)k2L2 (Ω) + kDukt (s)k2L2 (Ω) ds
0
h i Z t
≤ C ku0 kH 1 (Ω) + ku1 kH 1 (Ω) + kDukt (t)kL2 (Ω) + C kDukt (s)k2L2 (Ω) ds,
2 2 2
0
and
Z tZ
∂h ∂uktt
aij (s)
(s) (s)dxds
0 Ω ∂xj ∂xi
Z
∂h ∂ukt s=t Z t Z ∂ h ∂h i ∂u
ij kt
= a (s) (s) (s)dx − aij (s) (s) (s)dxds

Ω ∂x j ∂x i s=0 0 Ω ∂s ∂x j ∂x i
h i Z t
≤ C khkW 2 (T ;Ω)+ kukt (0)k2H 1 (Ω) + kDukt (t)k2L2 (Ω)+ C kDukt (s)k2L2 (Ω) ds.
0
Moreover, similar to (7.20) we have

Z t
hG, uκtt iH 0.5 (∂Ω) ds
0
h i
≤ C kG(0)k2H −0.5 (∂Ω) + ku1 k2H 1 (Ω) + kGt k2L2 (0,T ;H −0.5 (∂Ω))
Z t
2 2
+ C kG(t)kH −0.5 (∂Ω) + kukt (t)kH 1 (Ω) + C kukt (s)k2H 1 (Ω) ds
0
h i
≤ C ku1 kH 1 (Ω) + kGkL2 (0,T ;H −0.5 (∂Ω)) + kGt k2L2 (0,T ;H −0.5 (∂Ω))
2 2
Z th i
2
+ kDukt (t)kL2 (Ω) + C kuktt (s)k2L2 (Ω) + kDukt (s)k2L2 (Ω) ds;
0
thus integrating (7.115) in time over the time interval (0, t), by the positive
definiteness of aij and choosing > 0 small enough, we find that
h
kuktt (t)k2L2 (Ω) + kDukt (t)k2L2 (Ω) ≤ kuktt (0)k2L2 (Ω) + C ku0 k2H 1 (Ω) + ku1 k2H 1 (Ω)
+ kF k2L2 (0,T ;L2 (Ω)) + kGk2L2 (0,T ;H −0.5 (∂Ω)) + kGt k2L2 (0,T ;H −0.5 (∂Ω))
i Z th i
+ khkW 2 (T ;Ω) + C kuktt (s)k2L2 (Ω) + kDukt (s)k2L2 (Ω) ds.
0
Therefore, by the Gronwall inequality and that kuktt (0)kL2 (Ω) ≤ ku2 kL2 (Ω) ,
we conclude that
h
kuktt (t)kL2 (Ω)+ kukt (t)kH 1 (Ω) ≤ C ku0 kH 1 (Ω)+ ku1 kH 1 (Ω)+ ku2 kL2 (Ω)
i
+ kF kL2 (0,T ;L2 (Ω))+ kGkL2 (0,T ;H −0.5 (∂Ω))+ kGt kL2 (0,T ;H −0.5 (∂Ω))+ khkW 2 (T ;Ω)

for some constant C = C kakC 1 (Ω×[0,T ]) , kakC 1 (Ω×[0,T ]) , kakC 1 (Ω×[0,T ]) , λ 0 .
Similar to argument before, we also have
h
kukttt kL2 (0,T ;H 1 (Ω)0 ) ≤ C kukt kL2 (0,T ;H 1 (Ω)) + kuk kL2 (0,T ;H 1 (Ω))
i
+ kF kL2 (0,T ;L2 (Ω)) + kGkL2 (0,T ;H −0.5 (∂Ω)) + khkW 2 (T ;Ω)
h
≤ C ku0 kH 1 (Ω) + ku1 kH 1 (Ω) + ku2 kL2 (Ω) + kF kL2 (0,T ;L2 (Ω))
i
+ kGkL2 (0,T ;H −0.5 (∂Ω)) + kGt kL2 (0,T ;H −0.5 (∂Ω)) + khkW 2 (T ;Ω) .
Therefore, if u is a weak limit of uk in Lp (0, T ; H 1 (Ω)), then
kuttt kL2 (0,T ;H 1 (Ω)0 ) + kutt kL∞ (0,T ;L2 (Ω)) + kut kL∞ (0,T ;H 1 (Ω))
h
≤ C ku0 kH 1 (Ω) + ku1 kH 1 (Ω) + ku2 kL2 (Ω) + kF kL2 (0,T ;L2 (Ω)) (7.116)
i
+ kGkL2 (0,T ;H −0.5 (∂Ω)) + kGt kL2 (0,T ;H −0.5 (∂Ω)) + khkW 2 (T ;Ω) ,
and u satisfies that for all v ∈ L2 (0, T ; H 1 (Ω)),

Z Th

i
uttt (t), v(t) H 1 (Ω) + B1 u(t), v(t) + B2 u(t), v(t) dt
0
Z Th
∂h ∂ϕ i
Z
F (t), v(t) L2 (Ω) + G(t), v(t) H 0.5 (∂Ω) − aij (t)

= (t) dx dt
0 Ω ∂xj ∂xi
which implies that u satisfies (7.111) as well. The proof of showing u = u0

and ut = u1 at t = 0 in Section 7.4.1 can be applied to show the validity of
(7.112), so u is a weak solution to (7.110).
Lemma 7.25. Assume that aij , aij , aij 1 ij

k ∈ C (Ω × [0, T ]) satisfying a ξj ξi ≥
λ 0 |ξ|2 , u0 ∈ H 1 (Ω), u1 ∈ H 1 (Ω) and u2 ∈ L2 (Ω). Then for any given
F ∈ L2 (0, T ; L2 (Ω)), G, Gt ∈ L2 (0, T ; H −0.5 (∂Ω)), and hk ∈ W 2 (T ; Ω) for
k = 1, · · · , K, there exists a weak solution u ∈ L∞ (0, T ; H 1 (Ω)) to (7.110)
satisfying
h
≤ C ku0 kH 1 (Ω) + ku1 kH 1 (Ω) + ku2 kL2 (Ω) + kF kL2 (0,T ;L2 (Ω)) (7.117)
K
X i
+ kGkL2 (0,T ;H −0.5 (∂Ω)) + kGt kL2 (0,T ;H −0.5 (∂Ω)) + khk kW 2 (T ;Ω)
k=1

for some constant C = C kakC 1 (Ω×[0,T ]) , kakC 1 (Ω×[0,T ]) , kakC 1 (Ω×[0,T ]) , λ 0 .
Now let aij = aij

t , F = ft , G = gt and hk = 0 for all k in (7.110),
0.5 (T ).
and suppose that u0 ∈ H 2 (Ω), u1 ∈ H 1 (Ω), f ∈ FΩ0 (T ) and g ∈ G∂Ω
Integrating (7.111) in time over the time interval (0, t),
∂u0 ∂ϕ
Z
aij (x, 0)

utt (t), ϕ L2 (Ω) − (u2 , ϕ)L2 (Ω) + B1 u(t), ϕ − dx
∂xj ∂xi
Ω
= f (t), ϕ L2 (Ω) − f (0), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω) − g(0), ϕ L2 (∂Ω) .
Integrating by parts in xi , the equality above implies that

utt (t), ϕ L2 (Ω) + B1 u(t), ϕ = f (t), ϕ L2 (Ω) + g(t), ϕ L2 (∂Ω)
∂u0
+ u2 − aij (0)u0 ,j ,i −f (0), ϕ 2 + aij (0)

Ni − g(0), ϕ 2 .
L (Ω) ∂xj L (∂Ω)
Therefore, if u2 = f (0) + aij (0)u0 ,j ,i (which is (7.1a) evaluated at t = 0)

∂u0
and aij (0) Ni = g(0) (which is the first-order compatibility condition for
∂xj
(7.1a,b,c,e)), then u is the unique weak solution to (7.1a,b,c,e). Moreover,

by (7.117)
h
≤ C ku0 kH 1 (Ω) + ku1 kH 1 (Ω) + f (0) + aij (0)u0 ,j ,i L2 (Ω)

i
+ kft kL2 (0,T ;L2 (Ω)) + kgt kL2 (0,T ;H −0.5 (∂Ω)) + kgtt kL2 (0,T ;H −0.5 (∂Ω))
h
≤ C ku0 kH 2 (Ω) + ku1 kH 1 (Ω) + kft kL2 (0,T ;L2 (Ω))
i
+ kgt kL2 (0,T ;H −0.5 (∂Ω)) + kgtt kL2 (0,T ;H −0.5 (∂Ω))

for some constant C = C kakC 1 (Ω×[0,T ]) , kat kC 1 (Ω×[0,T ]) , λ 0 . The elliptic
regularity then implies that
h
kukL∞ (0,T ;H 2 (Ω)) ≤ C kf kL∞ (0,T ;L2 (Ω)) + kgkL∞ (0,T ;H 0.5 (∂Ω)) + kutt kL∞ (0,T ;L2 (Ω))
i
+ kukL∞ (0,T ;L2 (Ω))
h i
≤ C ku0 kH 2 (Ω) + ku1 kH 1 (Ω) + kf kF 0 (T ) + kgkG 0.5 (T ) .
Ω ∂Ω
If bi or c is non-zero, similar argument can be made; thus we establish the

following
Theorem 7.26. Assume that ∂Ω is of class C 2 , ∂tt + L is uniformly hy-

perbolic, aij , aij i 1 ij
0.5 (T ), the unique weak so-
u0 ∈ H 2 (Ω), u1 ∈ H 1 (Ω), f ∈ FΩ0 (T ), and g ∈ G∂Ω
lution u to the hyperbolic initial-boundary value problem (7.1a,b,c,e) in fact
belongs to L∞ (0, T ; H 2 (Ω)) provided that u0 satisfies
∂u0
aij (0) Ni = g(0) on ∂Ω . (7.118)
∂xj
Moreover, u satisfies
kukL∞ (0,T ;H 2 (Ω)) + kut kL∞ (0,T ;H 1 (Ω)) + kutt kL∞ (0,T ;L2 (Ω))
h i (7.119)
≤ C ku0 kH 2 (Ω) + ku1 kH 1 (Ω) + kf kF 0 (T ) + kgkG 0.5 (T )
Ω ∂Ω

for some constant C = C k(a, at , b, c)kC 1 (Ω×[0,T ])4 , λ 0 , T .
Now suppose further that u0 ∈ H 3 (Ω), u1 ∈ H 2 (Ω), f ∈ FΩ1 (T ) and

1.5 (T ), and aij ∈ C 3 (Ω × [0, T ]) (and bi = c = 0 for simplicity). By
g ∈ G∂Ω)
Theorem 7.26, if u0 satisfies the first-order compatibility condition (7.118),

then u ∈ W 2 (T ; Ω) satisfies (7.119).
Let w = ut . Then we may write the second time-differentiated version
of (7.110) as
∂ ij ∂wt ∂ ij ∂w ∂ ij ∂u
wttt − a − 2at = ftt + a in Ω × (0, T ) ,
∂xi ∂xj ∂xi ∂xj ∂xi tt ∂xj
∂wt ∂w ∂u
aij Ni + 2aij
t Ni = gtt − aij
tt Ni on ∂Ω × (0, T ) ,
∂xj ∂xj ∂xj
w = u1 , wt = u2 , wtt = u3 on Ω × {t = 0} ,
and the discussion above implies that the unique weak solution w satisfies
kwttt kL2 (0,T ;H 1 (Ω)0 ) + kwtt kL∞ (0,T ;L2 (Ω)) + kwt kL∞ (0,T ;H 1 (Ω))
h
≤ C ku1 kH 1 (Ω) + ku2 kH 1 (Ω) + ku3 kL2 (Ω) + kftt kL2 (0,T ;L2 (Ω))
i
+ kgtt kL2 (0,T ;H −0.5 (∂Ω)) + kgttt kL2 (0,T ;H −0.5 (∂Ω)) + kukW 2 (T ;Ω))
h i
≤ C ku1 kH 1 (Ω) + ku2 kH 1 (Ω) + ku3 kL2 (Ω) + kf kF 1 (T ) + kgkG 1.5 (T )
Ω ∂Ω

for some constant C = C kakC 2 (Ω×[0,T ]) , kat kC 2 (Ω×[0,T ]) , λ 0 . The second-
order compatibility condition
∂u1 ∂u0
Ni + aij
aij (x, 0) t (x, 0) Ni = gt (0)
∂xj ∂xj
and u3 = ft (0) + aij (0)u1 ,j ,i + aij

t (0)u0 ,j ),i then imply that u is the
unique weak solution to (7.110) (with aij = aij
t and h = 0). This observation
implies that the unique weak solution to (7.1a,b,c,e) (with bi = c = 0)
satisfies
kutttt kL2 (0,T ;H 1 (Ω)0 ) + kuttt kL∞ (0,T ;L2 (Ω)) + kutt kL∞ (0,T ;H 1 (Ω))
h i
Ω ∂Ω
Since u and ut satisfies
aij u,j ,i = f − utt

in Ω × (0, T ) ,
aij u,j Ni = g, on ∂Ω × (0, T ) ,
aij ut ,j ,i = ft − uttt − aij

t u,j ,i in Ω × (0, T ) ,
aij ut ,j Ni = gt − aij
t u,j Ni on ∂Ω × (0, T ) ,
by elliptic regularity,
h
kukL∞ (0,T ;H 3 (Ω)) ≤ C kf kL∞ (0,T ;H 1 (Ω))+ kutt kL∞ (0,T ;H 1 (Ω))+ kgkL∞ (0,T ;H 1.5 (∂Ω))
i
+ kukL∞ (0,T ;L2 (Ω))
h i
Ω ∂Ω
and
h
kut kL∞ (0,T ;H 2 (Ω)) ≤ C kft kL∞ (0,T ;L2 (Ω)) + kuttt kL∞ (0,T ;L2 (Ω)) +
i
+ kgt kL∞ (0,T ;H 0.5 (∂Ω)) + kukL∞ (0,T ;H 2 (Ω))
h i
≤ C ku0 kH 3 (Ω) + ku1 kH 2 (Ω) + kf kF 1 (T ) + kgkG 1.5 (T ) .
Ω ∂Ω
The combination of these estimates imply that u ∈ W 3 (T ; Ω) and satisfies

h i
kukW 3 (T ;Ω) ≤ C ku0 kH 3 (Ω) + ku1 kH 2 (Ω) + kf kF 1 (T ) + kgkG 1.5 (T ) . (7.120)
Ω ∂Ω
By induction, we establish the following
Theorem 7.27. Assume that ∂Ω is of class C m+1 , ∂tt + L is uniformly

hyperbolic, aij , aij i m ij
t , b , c ∈ C (Ω × [0, T ]), and [a ] is symmetric. Then for
m−0.5
any u0 ∈ H m+1 (Ω), u1 ∈ H m (Ω), f ∈ FΩm−1 (T ), and g ∈ G∂Ω (T ),
the unique weak solution u to the hyperbolic initial-boundary value problem
(7.1a,b,c,e) in fact belongs to W m+1 (T ; Ω) provided that u0 satisfies the
compatibility condition up to m-th order; that is,
k−1

ij
a (0) Ni = k (0) − Ni k−`
(0) on ∂Ω
`=0
for k = 1, · · · , m. Moreover, u satisfies

kukW m+1 (T ;Ω)
h i
≤ C ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T )
Ω ∂Ω
(7.121)

for some constant C = C k(a, at , b, c)kC m (Ω×[0,T ])4 , λ 0 , T .
7.5 Hyperbolic equations with Dirichlet boundary

condition
The proof of the existence and uniqueness of the weak solution to the hy-
perbolic initial-boundary value problem (7.1a,b,c,d) is essentially the same
7.6. EXERCISES 327
as for the Neumann case, and is left to the readers. Moreover, the regularity
theory for the Dirichlet problem is much easier than the Neumann case, be-
cause of the simpleness of the compatibility conditions (7.122). Therefore,
we state without proving the following
Theorem 7.28. Assume that ∂Ω is of class C m+1 , ∂tt + L is uniformly hy-

perbolic, aij , aij i m ij
u0 ∈ H m+1 (Ω), u1 ∈ H m (Ω), and f ∈ FΩm−1 (T ), the unique weak solution u
to the hyperbolic initial-boundary value problem (7.1a,b,c,d) in fact belongs
to W m+1 (T ; Ω) provided that u0 satisfies the compatibility condition up to
m-th order; that is,
τ uk = 0 on ∂Ω for k = 1, · · · , m. (7.122)
Moreover, u satisfies
h i
kukW m+1 (T ;Ω) ≤ C ku0 kH m+1 (Ω) + ku1 kH m (Ω) + kf kW m−1 (T ) + kgkW m−0.5 (T )
Ω ∂Ω

for some constant C = C k(a, at , b, c)kC m (Ω×[0,T ])4 , λ 0 , T .
7.6 Exercises
Problem 7.1. Mimic the proof of the existence of weak solutions to the
parabolic initial-boundary value problem (6.6a,b,d) to show that the weak
limit u obtained in (7.104) is in fact a weak solution to (7.1a,b,c,e). Note
that you need to show that u satisfies (7.6) with u(0) = u0 and ut (0) = u1 .
Also show that (7.6) and (7.105) are equivalent.
Problem 7.2. In this problem, we study the L∞ in time H 2 in space reg-

ularity theory of hyperbolic initial-boundary value problem (7.1a,b,c,e) with
g = 0 using a different approach. Complete the following.
1. We slightly modify Problem 5.9 in Chapter 5. Let aij ∈ C 1 (Ω) be

symmetric and positive definite satisfying
aij ξj ξi ≥ λ|ξ|2
for some constant λ > 0, and define an equivalent H 1 -norm by the

induced norm of the inner product
∂u ∂v
Z
((u, v))H 1 (Ω) ≡ (u, v)L2 (Ω) + aij dx .
Ω ∂xj ∂xi
For a given f ∈ L2 (Ω), let Kf be the unique solution to the (elliptic)

Neumann problem
∂ ij ∂u
u− a =f in Ω,
∂xi ∂xj
∂u
aij Ni = 0 on ∂Ω .
∂xj
Show that K : L2 (Ω) → H 2 (Ω) is bounded, and K : L2 (Ω) → L2 (Ω) is

symmetric. Construct an orthonormal basis in L2 (Ω) which is orthog-
onal in the product ((·, ·))H 1 (Ω) .
k
Let {e` }∞
P
`=1 be such a basis, and uk = (u, e` )L2 (Ω) e` . Show that
`=1
uk → u in H 1 (Ω) for all u ∈ H 1 (Ω), and
kuk kH 1 (Ω) ≤ CkukH 1 (Ω) ∀ u ∈ H 1 (Ω) (7.123)
∂u
for some C > 0. Moreover, if u ∈ H 2 (Ω) satisfies aij Ni = 0 on
∂xj
∂Ω, show that uk → u in H 2 (Ω), and for some constant C > 0,
∂u
kuk kH 2 (Ω) ≤ CkukH 2 (Ω) ∀ u ∈ H 2 (Ω) with aij Ni = 0 . (7.124)
∂xj
2. Let aij (x) = aij (x, 0) in 1. Use {e` }∞ 2

`=1 as the L -orthonormal ba-
sis in the Galerkin scheme to construct the unique weak solution: let
k
dkm (t)em (x) be the finite dimensional approximation of
P
uk (x, t) =
m=1
the solution, where dk satisfies ODE (7.96) (with M and fˆ defined
using the new basis). Show that
kuktt (t)k2L2 (Ω) + λ 0 kDukt (t)k2L2 (Ω)

h i
≤ kuktt (0)k2L2 (Ω) + C ku1 k2H 1 (Ω) + ku0 k2H 1 (Ω) (7.125)
Z th i
+C kuktt (s)k2L2 (Ω) + kDukt (s)k2L2 (Ω) ds
0

for some constant C = C kakC 1 (Ω) , kat kC 1 (Ω×[0,T ]) .
§7.6 Exercise 329
3. Show that
kuktt (0)kL2 (Ω) ≤ kvkL2 (Ω) ≤ kf (0)kL2 (Ω) + Ckuk (0)kH 2 (Ω)
h i
≤ C ku0 kH 2 (Ω) + kf kF 0 (T ) (7.126)
Ω
k
using the fact that kuktt (0)k2L2 (Ω) = |dkm00 (0)|2 .
P
m=1
Hint: You can start from the following equality
k
= fˆm (0) −
X
dkm00 (0) M`m (0)dk` (0)
Z`=1h
∂uk i
= fˆm (0) − bi (x, 0) (x, 0)em (x) + c(x, 0)uk (x, 0)em (x) dx
Ω ∂xi
∂u ∂em
Z
k
− aij (x, 0) (x, 0) (x) dx .
Ω ∂xj ∂xi
4. What can you say about the case g 6= 0 using this approach?
Problem 7.3. Prove the existence and uniqueness of the solution of (7.29)
in L2 (0, T ; H m+1 (Ω)) using the fixed-point argument.
Problem 7.4. Prove Theorem 7.23.
Problem 7.5. Prove Theorem 7.28 using
1. the method of parabolic regularization;
2. the Galerkin method.
Problem 7.6. Let Ω ⊂ R3 denote an open, bounded, connected and smooth

domain. Prove existence and uniqueness of weak solutions to
utt + Dp = 0 in Ω × (0, T ] ,
div u = 0 on Ω × (0, T ] ,
pN = −∆g u on ∂Ω × [0, T ] ,
(u, ut ) = (u0 , u1 ) on Ω × {t = 0} ,
where ∆g denotes the Laplace-Beltrami operator on ∂Ω defined in (5.6), and

N denotes the outward unit-normal to ∂Ω. To do so, consider an approxi-
mation given by the parabolic κ-problem
uκtt + Dpκ = 0 in Ω × (0, T ] ,

div uκ = 0 on Ω × (0, T ] ,
pκ N = −∆g uκ − κ∆g uκt on ∂Ω × [0, T ] ,
(u, ut ) = (u0 , u1 ) on Ω × {t = 0} .
Use a contraction mapping algorithm to establish solutions to the κ-problem,

establish κ-independent estimates, and find the limit as κ → 0. Write the
basic inequality satisfied by the weak solution.
Problem 7.7. Consider the wave equation
utt − Ddivu = 0 in T3 × (0, T ] ,

(u, ut ) = (u0 , u1 ) on T3 × {t = 0} .
1. Prove existence and uniqueness of weak solutions;
2. prove u ∈ L∞ (0, T ; H 1 (T3 )) and write the basic inequality that u sat-
isfies.
(Hint. Consider the parabolic κ-problem
utt − κD div ut − D div u = 0 in T3 × (0, T ] ,

(u, ut ) = (u0 , u1 ) on T3 × {t = 0} .)
Problem 7.8. Let Ω ⊆ R3 denote an open, bounded, connected and smooth

domain. Consider the semi-linear Sine-Gordon wave equation:
utt − ∆u + sin(u) = 0 in Ω × (0, T ] ,

u=0 on ∂Ω × (0, T ] ,
(u, ut ) = (u0 , u1 ) on ∂Ω × {t = 0} .
Use the contraction mapping theorem to find the unique weak solution to the
Sine-Gordon equation. (Hint. Consider the parabolic κ-problem
utt − κ∆ − ∆u + sin(u) = 0 in Ω × (0, T ] ,

u=0 on ∂Ω × (0, T ] ,
(u, ut ) = (u0 , u1 ) on ∂Ω × {t = 0} .)
§7.6 Exercise 331
State the compatibility conditions for the initial data which are required to
obtain a solution u ∈ L∞ (0, T ; H 2 (Ω)).
Chapter 8
More Topics on Elliptic

Equations
8.1 Elliptic equations with Sobolev class coeffi-

cients
We continue to assume that Ω ⊂ Rn is a bounded, smooth, and open subset,

and once again consider the second-order elliptic equation
n
X ∂ ij ∂u
− a =f in Ω, (8.1a)
∂xj ∂xi
i,j=1
u=0 on ∂Ω , (8.1b)
under the assumption that aij ∈ H k (Ω) (rather than C k (Ω)) for k > n/2,
and satisfies
n
X
0 < λ0 |ξ|2 ≤ aij (x)ξi ξj ∀ ξ ∈ Rn − {0} (8.2)
i,j=1
for some positive constant λ0 . The Lax-Milgram theorem then guarantees

the existence of a weak solution u ∈ H01 (Ω), with kukH01 (Ω) ≤ Ckf kL2 (Ω) . It
remains to develop the regularity theory and establish new estimates for the
solution which depend crucially on kakH k (Ω) .
n
Theorem 8.1. For k > , f ∈ H k−1 (Ω), and aij ∈ H k (Ω) satisfying (8.2),
2
the weak solution u ∈ H01 (Ω) to (8.1) is an H k+1 (Ω)-strong solution, and
h i
kukH k+1 (Ω) ≤ C kf kH k−1 (Ω) + P(kakH k (Ω) )kf kL2 (Ω) . (8.3)
332
8.1. ELLIPTIC EQUATIONS WITH SOBOLEV CLASS COEFFICIENTS333
Proof. Let aij ij

= η ∗ E(a ), where η is the standard mollifier defined in
Definition 2.25, and E : H k (Ω) → H k (Rn ) is the Sobolev extension operator.

For > 0 taken sufficiently small, aij
is strictly elliptic, and
n
X
2
λ0 |ξ| ≤ aij
(x)ξi ξj ∀ ξ ∈ Rn .
i,j=1
Moreover, aij ij k
→ a in H (Ω) as → 0 . By elliptic regularity, the solution
u to
n
X ∂ ij ∂u
− a (x) (x) = (η ∗ f )(x) in Ω, (8.4a)
∂xj ∂xi
i,j=1
u = 0 on ∂Ω , (8.4b)
belongs to H ` (Ω) for all ` ∈ N . We show that u has a uniform bound in

H k+1 (Ω) and the weak limit of u in H k+1 (Ω) is indeed the solution to (8.1).
Let {χm }N
m=0 be a partition of unity subordinate to an open cover
{Um }N
m=0 of Ω, such that Um ∩ ∂Ω 6= ∅ for all m ≥ 1 , U0 ∩ ∂Ω = ∅ , and
there exist smooth H k -class diffeomorphisms ψm : Vm ≡ B(0, 1) ∩ Rn+ → Um
for all m ≥ 1 . (The maps ψm can be taken smooth whenever Ω is smooth.)
Step 1. Boundary regularity Let χ = χm , ψ = ψm , U = Um , and
V = Vm for some fixed m ≥ 1 . Multiplying the both sides of (8.4a) by χ ,
and then composing both sides with ψ , we have
n
h X ∂ ij ∂u i
− χ a ◦ ψ = (χη ∗ f ) ◦ ψ in V .
∂xj ∂xi
i,j=1
Let yk denote the local coordinates on V , and let v = u ◦ ψ . Then the

equation above reads
n
X ∂ ij ∂v h i
−χ̃(y) Akj (y) ã (y)Aì (y) (y) = (χη ∗ f ) ◦ ψ (y) ,
∂yk ∂y`
i,j,k,`=1
where χ̃ = χ ◦ ψ , ãij ij
= a ◦ ψ , and A = (Dψ)
−1 . Let J = det(Dψ) . By
Piola’s identity,
n
X ∂ ∂v
− χ̃(y) J(y)Akj (y)Aì (y)ãij
(y) (y)
∂yk ∂y`
i,j,k,`=1 (8.5)
h i
= J(χη ∗ f ) ◦ ψ (y) ≡ F (y) ∀y ∈ V .
334 CHAPTER 8. MORE TOPICS ON ELLIPTIC EQUATIONS
n
Let bk` = JAkj Aì ãij
P
. Since ψ : V → U is diffeomorphism, J may be
i,j=1
assumed to have a positive lower bound, say J ≥ J0 > 0 . Then
n
X n
X
bk` ξk ξ` = Jãij k ` T 2 2
(Ai ξk )(Aj ξ` ) ≥ J0 λ0 |A ξ| ≥ J0 λ0 c0 |ξ| , (8.6)
k,`=1 i,j,k,`=1
where c0 is the smallest eigenvalue of AAT which has a positive lower bound.
In other words, bk` is also strictly elliptic.
Letting w = χ̃v , we rewrite (8.5) as
n n h
X ∂ ij ∂w X ∂ χ̃ ∂v ∂(bij χ̃,j ) i
− b =F− (bij + bji ) + v .
∂yi ∂yj ∂yi ∂yj ∂yi
i,j=1 i,j=1
Letting ∂¯α (the α-th tangential differentiation operator) act on the equation
above, we find that ∂¯α w satisfies
n h i n h i
b (∂¯α w ),j = ∂¯α F + Cαβ (∂¯β bij )(∂¯α−β w ),j
X X X
ij
−
,i ,i
i,j=1 i,j=1 0<β≤α
n h i
Cαβ ∂¯β (bij + bji )χ̃,i ∂¯α−β v,j + ∂¯β (bij χ̃,j ),i ∂¯α−β v
X X
−
i,j=1 0≤β≤α
(8.7)
with the boundary condition ∂¯α w = 0 on ∂V . Note that since v ∈ H ` (Ω)
for all ` ∈ N , ∂¯α w is a classical solution to the equation above; hence a
weak solution as well. Therefore,
h
∂¯ w ≤ C ∂¯α F H 1 (V)0
α
H 1 (V)

∂¯j Db∂¯|α|−` v 1 0 + ∂¯j b∂¯|α|−` v 1 0 + ∂¯j b∂¯|α|−` Dv 1 0
X
+ H (V) H (V) H (V)
0≤j≤`≤|α|
|α| |α|−`
∂¯` Db∂¯|α|−`−j v 1 0 + ∂¯` Db∂¯|α|−`−j Dv 1 0
X X
+ H (V) H (V)
`=1 j=0
i
+ ∂¯` b∂¯|α|−`−j v H 1 (V)0 + ∂¯` b∂¯|α|−`−j Dv H 1 (V)0 + ∂¯` b∂¯|α|−`−j D2 v H 1 (V)0 .

Let σ(n) = 0 if n ≥ 3 or σ(n) = 0.5 if n = 2 . Since
h i
¯ kH 1 (V)0 ≤ Cn kf gkL2 (V) + kf ∂gk
kg ∂f ¯ 2n ,
L n+2−σ(n) (V)
8.1. ELLIPTIC EQUATIONS WITH SOBOLEV CLASS COEFFICIENTS335
we find that for |α| ≤ k ,

h i
Dv ∂¯ Db 1 0 ≤ Ck∂¯k−1 DbkL2 (V) kDv kL∞ (V) + k∂Dv ¯ k 2n
|α|
H (V) L 2−σ (V)
h n 2k−n n+σ 2k−n−σ i
≤ Cn kbkH k (V) kv kH
2k 2k
k+1 (V) kv kH 1 (V) + kv kH k+1 (V) kv kH 1 (V)
2k 2k
h 2k 2k i
(Young’s inequality) ≤ Cnδ kbk k
2k−n
H (V)
+ kbk 2k−n−σ
H k (V)
kv kH 1 (V) + δkv kH k+1 (Ω)
≤ Cδ P(kakH k (Ω) )kf kL2 (Ω) + δkv kH k+1 (V) ,
where P denotes a polynomial function. Similarly, for 1 ≤ ` ≤ k − 1 ,
∂¯ Db∂¯k−` Dv 1 0 ≤ Cδ P kakH k (Ω) kf kL2 (Ω) + δkv kH k+1 (V) .

`
H (V)
Therefore, since k∂¯α F kH 1 (V)0 ≤ CkF kH |α|−1 (V) ≤ Ckf kH |α|−1 (Ω) ,
k∂¯α w kH 1 (V) ≤ Ckf kH k−1 (Ω) + Cδ P(kakH k (Ω) )kf kL2 (Ω) + δkv kH k+1 (V) .
X
|α|≤k
By (8.7),
− bnn ∂¯α w,nn

X h i n h i
ij ¯α ¯ Cαβ (∂¯β bij )(∂¯α−β w ),j
X X
α
= b (∂ w ),j + ∂ F +
,i ,i
1 ≤ i, j ≤ n i,j=1 0<β≤α
(i, j) 6= (n, n)
(8.8)
n h i
Cαβ ∂¯β (bij + bji )χ̃,i ∂¯α−β v,j + ∂¯β (bij χ̃,j ),i ∂¯α−β v .
X X
−
i,j=1 0≤β≤α
The right-hand side consists of terms with (|α| + 2)-th derivatives in which
only one derivative with respect to yn is involved (the first term) or terms
with lower order mixed derivatives (all the other terms). Hence, as bnn > 0,
it follows that
X
|α|≤k−1
Differentiating (8.8) with respect to yn then shows that
X
|α|≤k−2
A standard induction argument provides the estimate
kw kH k+1 (V) ≤ Ckf kH k−1 (Ω) + Cδ P(kakH k (Ω) )kf kL2 (Ω) + δkv kH k+1 (V) .
which, in turn, shows that
kχm u kH k+1 (Um ) ≤ Ckf kH k−1 (Ω) + Cδ P(kakH k (Ω) )kf kL2 (Ω) + δku kH k+1 (Ω) .
(8.9)
Step 2. Interior regularity Multiplying the both sides of (8.4a) by χ0 ,
we find that in U0 ,
n n h
X ∂ ij ∂(χ0 u ) X
ji ∂χ0 ∂u
∂(aij
χ0,j )
i
− a = χη ∗ f − (aij
+ a ) + u .
∂xj ∂xi ∂xi ∂xj ∂xi
i,j=1 i,j=1
Let Dα act on the equation above, since Dα w = 0 on ∂Ω , exactly the same

procedure as in the boundary regularity gives us
kχ0 u kH k+1 (U0 ) ≤ Ckf kH k−1 (Ω) + Cδ P(kakH k (Ω) )kf kL2 (Ω) + δku kH k+1 (U0 ) .
(8.10)
Step 3. Global regularity Combining (8.9) and (8.10), and taking δ > 0
sufficiently small, we obtain the -independent bound:
h i
ku kH k+1 (Ω) ≤ C kf kH k−1 (Ω) + P(kakH k (Ω) )kf kL2 (Ω) . (8.11)
0
Passing to the limit as 0 → 0 for a subsequence 0 we see that u → u in
H k+1−ς (Ω) for ς > 0, and that hence u is the unique solution of (8.1), and
satisfies the estimate (8.11).
8.2 Elliptic decomposition using the divergence and

curl
8.2.1 An H 1 (Ω) elliptic estimate
Let u : Ω ⊆ R3 → R3 be a smooth vector field. Then
curlcurlu = −∆u + Ddivu or − ∆u = curlcurlu − Ddivu . (8.12)
Testing (8.12) against u and integrating by parts, we find that

X3 Z Z
2 i i 2
kDukL2 (Ω) − u,j u Nj dS = kdivukL2 (Ω) − divu(u · N)dS
i,j=1 ∂Ω ∂Ω
3
X Z
+ kcurluk2L2 (Ω) − εijk εkrs us,r ui Nj dS ,
i,j,k,r,s=1 ∂Ω
8.2. ELLIPTIC DECOMPOSITION USING THE DIVERGENCE AND CURL337
where εijk = 1(−1) if i, j, k is an even (odd) permutation of 1, 2, 3 . Since

P3
εijk εkrs = δir δks − δis δkr ,
k=1
Z 3
hX i
kDuk2L2 (Ω) = kdivuk2L2 (Ω) +kcurluk2L2 (Ω) + ui,j uj Ni −divu(u·N) dS .
∂Ω i,j=1
We first assume that u · N = 0 . Then

3
X 3
X X2 3 X
X 2
ui,j uj Ni = Nj ∂N ui + τjα ui,α uj Ni = − ui (uj τjα )Ni,α .
i,j=1 i,j=1 α=1 i,j=1 α=1
The equality above implies

3 X
X 2 Z
kDuk2L2 (Ω) = kdivuk2L2 (Ω) + kcurluk2L2 (Ω) − ui (uj τjα )Ni,α dS
i,j=1 α=1 ∂Ω
≤ kdivuk2L2 (Ω) + kcurluk2L2 (Ω) + C kuk2L2 (Ω) + kuk2H 1 (Ω)
or
h i
kuk2H 1 (Ω) ≤ C kuk2L2 (Ω) + kdivuk2L2 (Ω) + kcurluk2L2 (Ω) , (8.13)
where C = C(|∂Ω|H 2.5 ) depends on the regularity of ∂Ω . By a density
argument, (8.13) holds for all u ∈ H 1 (Ω) with u · N = 0 on ∂Ω .
Remark 8.2. If u = 0 on ∂Ω , then the constant C is independent of the

regularity of ∂Ω . In fact, for the case u = 0 on ∂Ω , we have
kDuk2L2 (Ω) = kdivuk2L2 (Ω) + kcurluk2L2 (Ω) .
Now suppose that u · N 6= 0 . By elliptic regularity, the solution ϕ to the

problem
1 ∂ϕ
Z
∆ϕ = u · N dS in Ω , =u·N on ∂Ω
|Ω| ∂Ω ∂N
satisfies
kϕk2H 2 (Ω) ≤ Cku · Nk2H 0.5 (∂Ω) .
Let v = u − Dϕ ∈ H 1 (Ω) , then v · N = 0 and hence
h i
kvk2H 1 (Ω) ≤ C kvk2L2 (Ω) + kdivvk2L2 (Ω) + kcurlvk2L2 (Ω)
which implies
h
kuk2H 1 (Ω) ≤ C(|∂Ω|H 2.5 ) kuk2L2 (Ω) + kdivuk2L2 (Ω) + kcurluk2L2 (Ω) (8.14)
i
+ ku · Nk2H 0.5 (∂Ω) ∀ u ∈ H 1 (Ω) .
8.2.2 An H 2 (Ω) elliptic estimate
Suppose that Ω is an H 3 class domain (with H 2.5 -class boundary ∂Ω), and
u ∈ H 2 (Ω) . Let χ` ∈ C0∞ (Rn ) denote a partition of unity subordinate
to the open cover U` , ` = 0, 1, · · · , N of Ω, such that U0 ∩ ∂Ω = ∅ and
U` ∩∂Ω 6= ∅ for all ` 6= 0 , and for each ` 6= 0 , there exist H 3 -diffeomorphisms
ϕ` : B(0, 1) ∩ Rn+ → U` ∩ Ω . Since D(χ0 u) ∈ H 1 (Ω) and ∂(χ0 u) · N = 0 on
∂Ω , (8.14) implies that
h i
kD(χ0 u)k2H 1 (Ω) ≤ C kuk2H 1 (Ω) + kdivuk2H 2 (Ω) + kcurluk2H 2 (Ω) .
Next, let χ = χ` and ψ = ψ` = ϕ−1 for some fixed ` 6= 0 . Let A =

(Dψ)−1 and v = u ◦ ψ and η = χ ◦ ψ. Then, locally, on B(0, 1) ∩ Rn+ , we see
that
3 3 h i 3
Aji (ηv)j,iα = Aji (ηv)i,j (Aji ),α (ηv i ),j ,
X X X
[div∂α (χu)] ◦ ψ = −
,α
i,j=1 i,j=1 i,j=1
3
X 3 h
X i X3
[curl∂α (χu)]i ◦ ψ = εijk A`j (ηv)k,`α = εijk A`j (ηv)k,` − εijk (A`j ),α (ηv)k,` .
,α
j,k,`=1 j,k,`=1 j,k,`=1
Therefore,
h i
kdiv∂α (χu)kL2 (Ω) ≤ C(|∂Ω|2.5 ) kdiv(ηu)kH 1 (Ω) + k∂AkL4 (Ω) kηvkW 1,4 (Ω)
h i
≤ C(|∂Ω|2.5 ) kdivukH 1 (Ω) + kuk 47 .
H (Ω)
Similarly,
h i
kcurl∂α (χu)kL2 (Ω) ≤ C(|∂Ω|2.5 ) kcurlukH 1 (Ω) + kuk 7 .
H 4 (Ω)
Since ∂α (χu) ∈ H 1 (Ω) , by (8.14),

h
k∂α (χu)k2H 1 (Ω) ≤ C(|∂Ω|2.5 ) k∂α (χu)k2L2 (Ω) + kdivuk2H 1 (Ω) + kcurluk2H 1 (Ω)
i
+ kuk2 7 + k∂α (χu) · Nk2H 0.5 (∂Ω) .
H 4 (Ω)
Together with (8.12), we find that

h
kuk2H 2 (Ω) ≤ C(|∂Ω|2.5 ) kuk2L2 (Ω) + kdivuk2H 1 (Ω) + kcurluk2H 1 (Ω)
i
¯ · Nk2 0.5
+ k∂u H (∂Ω) ,
8.3. THE PROBLEM OF DIVU = F IN Ω WITH U = 0 ON ∂Ω 339
¯ denotes the tangential derivative of u along ∂Ω. By induction, in

where ∂u
general we have for k ≥ 1 ,
h
kuk2H k+1 (Ω) ≤ C(|∂Ω|H ` ) kuk2L2 (Ω) + kdivuk2H k (Ω) + kcurluk2H k (Ω)
i
¯ · Nk2 k−0.5
+ k∂u ∀ u ∈ H k+1 (Ω) ,
H (∂Ω)
(8.15)
where ` = max{2.5, k + 0.5} .
If ∂Ω possesses better regularity, say ∂Ω ∈ H ` with ` = max{3.5, k+1.5} ,
then by ∂u · N = ∂(u · N) − u · ∂N ,
h
i
+ ku · Nk2H k+0.5 (∂Ω) ∀ u ∈ H k+1 (Ω) .
(8.16)
8.3 The problem of divu = f in Ω with u = 0 on ∂Ω
In this section, we discuss a very important problem in the study of fluid

dynamics: given f ∈ L2 (Ω)/R (i.e., Ω f (x)dx = 0), find a vector field u so
R
that divu = f in Ω and u = 0 on ∂Ω .

This problem is difficult because we not only require that u · N vanishes
on ∂Ω but also that u · τα vanishes, where τ1 and τ2 span the tangent space.
If only u · N = 0 on the boundary, then the solution is very easy to find: let
p be a solution to
−∆p = f in Ω,
∂p
=0 on ∂Ω ,
∂N
and u = Dp is a solution to divu = f in Ω with u · N = 0 on ∂Ω . The
argument for obtaining a solution u ∈ H01 (Ω) is non-trivial, and we have the
following theorem.
Theorem 8.3. Let n ≥ 2 and Ω ⊂ Rn a bounded,

Z open domain with C 3 -
boundary. Then for all f ∈ L2 (Ω) so that f (x)ds = 0 , there is a u ∈
Ω
H01 (Ω) satisfying divu = f and
kukH 1 (Ω) ≤ Ckf kL2 (Ω) (8.17)
for some constant C depending on |∂Ω|W 3,∞ .

Proof. We will provide the proof for the case n = 3. When n 6= 3, a slight
modification of this argument using differential forms is necessary.
Step 1. Existence. We look for a solution in the form
u = curlv + Dp .
Let p be denote a solution to the Neumann problem
−∆p = f in Ω, (8.18a)
∂p
=0 on ∂Ω . (8.18b)
∂N
∂p 1
We have that kpkH 2 (Ω) ≤ Ckf kL2 (Ω) and hence ∂N ∈ H 2 (∂Ω).
We must find a vector-field v ∈ H 2 (Ω) such that
curlv + Dp = 0 on ∂Ω .
1
Let g := N × Dp ∈ H 2 (∂Ω), and define B to be the unique H 2 (Ω)-
solution of
1
Z
−∆B = gdS in Ω,
|Ω| ∂Ω
∂B
=g on ∂Ω .
∂N
Let d(x) = inf |x − y| denote the distance function to the boundary. Since
y∈∂Ω
∂Ω is C 3 , d is three times differentiable near the boundary and −Dd = N
on the boundary. Let ξ : [0, ∞) → [0, ∞) be a smooth function with (small)
compact support so that ξ(0) = 1 and ξ 0 (0) = 0 . We claim that

v(x) = B(x) − ξ d(x) B x − d(x)Dd(x)
satisfies curlv = −Dp on ∂Ω .

To see this, we differentiate and obtain
v,jk (x) = B,jk (x) − ξ 0 d(x) d,j (x)B k x − d(x)Dd(x)

3
X h i
ξ d(x) B,`k x − d(x)Dd(x) δj` − d,j (x)d,` (x) − d(x)d,j` (x) .

−
`=1
For x ∈ ∂Ω , d(x) = 0 . Since ξ 0 (0) = 0 , we find that

3
X ∂B k
v,jk (x) = B,`k (x)d,j (x)d,` (x) = Nj = g k Nj ∀ x ∈ ∂Ω .
∂N
`=1
Therefore, on the boundary ∂Ω ,

3
X 3
X 3
X
i
(curlv) = εijk v,jk = k
εijk g Nj = εijk εkrs p,s Nr Nj
j,k=1 j,k=1 j,k,r,s=1
3
X ∂p
= (δir δjs − δis δjr )p,s Nr Nj = Ni − p,i N · N = −p,i ,
∂N
j,r,s=1
∂p
where the condition = 0 is used to conclude the last equality.
∂N
Step 2. Estimate. First we note that by elliptic regularity B satisfies
kBk2H 2 (Ω) ≤ C(|∂Ω|W 2,∞ )|g|2H 0.5 (Γ) ≤ C(|∂Ω|W 2,∞ )kpk2H 2 (Ω)
≤ C(|∂Ω|W 2,∞ )kf k2H 1 (Ω)
and hence v satisfies
kvk2H 2 (Ω) ≤ C(|∂Ω|W 3,∞ )kBk2H 2 (Ω) ≤ C(|∂Ω|W 3,∞ )kf k2H 1 (Ω) ,
where the dependence of |∂Ω|W 3,∞ comes from the estimate of D3 d . This
estimate, together with the estimate for p, leads to (8.17).
q
Remark Z 8.4. In general, if ∂Ω is locally Lipschitz, then for all f ∈ L (Ω)
so that f (x)dx = 0 , there exists u ∈ W01,q (Ω) satisfying divu = f and
Ω
kuk2W 1,q (Ω) ≤ Ckf k2Lq (Ω) ,
where C depends on Ω .
Corollary 8.5. Let Ω be a domain 3 Then for all f ∈

Z with C -boundary.
Z
2 0.5
L (Ω) and g ∈ H (Γ) so that f (x)ds = g(x) · N(x)dS , there is a
Ω Γ
u ∈ H 1 (Ω) so that divu = f in Ω and u = g on Γ with the estimate
h i
kuk2H 1 (Ω) ≤ C kf k2L2 (Ω) + kgk2H 0.5 (∂Ω)
for some constant C depending on |∂Ω|W 3,∞ .
8.3.1 Inequalities associated with the normal and tangential

decomposition of vector fields on ∂Ω
The regularity of ∂Ω
Definition 8.6. A (bounded) domain Ω ⊆ Rn is said to be of class C k if

∂Ω is an n − 1 dimensional C k sub-manifold of Rn , or equivalently,
N
(1) there are open sets O` ⊂ Rn such that ∂Ω ⊆
S
O` ;
`=1
(2) there are maps ϕ` : U` := ∂Ω ∩ O` → Rn−1 so that ψ` := ϕ−1

` :
ϕ` (U` ) → Rn is a C k injective immersion, that is, for each `, ψ` is
C k , one-to-one, and for a given coordinate (y 1 , · · · , y n−1 ) in ϕ` (U` ),
the set of vectors {ψ`,1 (y), · · · , ψ`,n−1 (y)} are linearly independent for
all y ∈ ϕ` (U` );
(3) and each transition map ϕ`1 ◦ ϕ−1

`2 : U`1 ∩ U`2 → R
n−1 is C r for some
r ≥ k.
A domain Ω is smooth if Ω is a C k -domain for all k > 0.
Given a local chart (U, ϕ) on ∂Ω at a point y ∈ ∂Ω with local coordinates

yα, α = 1, ..., n − 1, let Ty ∂Ω denote the tangent space to ∂Ω at the point y:
Ty ∂Ω = {v ∈ Rn |v · N (y) = 0} .
Each such tangent space is diffeomorphic to Rn−1 with n − 1 linearly inde-

pendent tangent vectors spanning the space. The natural basis of tangent
vectors is given by the partial derivatives of ψ:
∂
:= ψ,α (y) α = 1, ..., n − 1 ,
∂y α y
which condition (2) in Definition 8.6, are linearly independent.

For a function f defined on ∂Ω, the partial derivative of f with respect
to y α at y, denoted by f,α (y), is defined as
∂f ∂(f ◦ ψ)
f,α (y) = α
(y) := . (8.19)
∂y ∂y α
Throughout, Greek indices run from 1 to n − 1, while Latin indices run from
1 to n.
Having established a local coordinate system y α , α = 1, ..., n − 1 on
the boundary ∂Ω, we extend the coordinate system to an n-dimensional
neighborhood of ∂Ω. For y n ∈ R, define the map Ψ : ϕ(U) × R → Rn by
Ψ(y 1 , · · · , y n−1 , y n ) = ψ(y 1 , · · · , y n−1 ) + y n N (ψ(y 1 , · · · , y n−1 )) , (8.20)

where N denotes the outward pointing unit normal on ∂Ω. With the diame-
ter of O taken sufficiently small, Ψ : V = Φ(O) → O is bijection mapping the
curvilinear cooridantes (y1 , ..., yn ) onto the standard Cartesian coordinates
(x1 , ..., xn ).
Ψ Ω
Φ(O+ ) ∂Ω
ψ = ϕ−1
O+
(y1 , · · · , yn−1 ) ∈ "n−1
ϕ
y ∈ ∂Ω
Φ = Ψ−1
Let p ∈ O+ = O ∩Ω. By assumption there exists a unique q = q(p) ∈ ∂Ω

so that dist(p, q) = dist(p, ∂Ω), and
*
pq = dist(p, q)N (q) .
Suppose p = Ψ(y 1 , · · · , y n ), then q = Ψ(y 1 , · · · , y n−1 , 0). Given y ∈ O+ , we

define the tangent vectors
∂
:= Ψ,i (y) .
∂y i y
Next, we define
n := N ◦ ψ .
It follows that
∂ h i
n
:= ψ,α + y n ,α (y 1 , · · · , y n−1 ) ; (8.21a)
∂y α y
∂
:= n(y 1 , · · · , y n−1 ) . (8.21b)
∂y n y
N
S
By adding an open set O0 ⊂⊂ Ω, we may assume that Ω ⊆ O` . Let
`=0
{χ` }N
`=0 be a partition of unity (of Ω) subordinate to O` . We will further
assume that spt(χ` ) ⊂⊂ O` , O` is smooth for all `, and ∂O` is transverse to
{yn = 0} (and equivalently, ∂Φ(O` ) is transverse to ∂Ω).
Let f be a function defined on Ω, and f` = χ` f . By the properties of
N N
{χ` }N
P P
`=0 , f = (χ` f ) = f` . For 1 ≤ i ≤ n, f` ◦ Ψ` is defined on Φ(O` ),
`=0 `=0
or in other words, f` ◦ Ψ` is a function of y. Let A` = (Dy Ψ` )−1 , by the

chain rule,
∂f` ∂(f` ◦ Ψ` )
◦ Ψ` = (A` )ji = (A` )αi f`,α + (A` )ni f`,n . (8.22)
∂xi ∂yj
Important notation: For a given function f , we use f,i to denote the
derivative of f with respect to the Cartesian coordinates xk , and we use
f,α and f,n to denote the derivative of f with respect to the curvilinear
coordinates yk .
Tangential and normal derivatives
Decomposing vectors on ∂Ω in 2-D. Suppose that n = 2. We begin with the

definition of two important geometric quantities: in each chart (U, ϕ), let
g = ψ,1 ·ψ,1 and b = n,1 ·ψ,1 denote the first and second fundamental forms,
respectively. Then
ψ 2 ,1 ψ 1 ,1
n = (− √ , √ ) , and n,1 = g −1 bψ,1 .
g g
By the definition of Ψ,
2 
ψ,1 2 
ψ,1
y2 b
 
1 1 1
 ψ,1 + y2 n,1 − √g   ψ,1 (1 + g
) −√ 
g 
Dy Ψ =  = ;
  
1
ψ,1   2 1
ψ,1

2 2 y2 b 
ψ,1 + y2 n,1 √ ψ,1 (1 + ) √
g g g
g + y2 b
hence, det(Dy Ψ) = √ and
g
 1
ψ,1 2
ψ,1

√ √ √
g  g g 
A=  .
g + y2 b 
2 y2 b 1 y2 b 
−ψ,1 (1 + ) ψ,1 (1 + )
g g
Consequently,
1
τi1 := A1i = ψ,1 // ∂Ω , and A2i = ni .
g + y2 b
and (8.22) implies that the gradient can be decomposed into tangential and
normal derivatives:
∂f
◦ Ψ = τi1 f,1 + ni f,n ,
∂xi
where f,1 = ∂f /∂y1 and f,n = ∂f /∂y2 .

Decomposing vectors on ∂Ω in n-D. Let eα = (0, · · · , 0, 1, 0, · · · , 0)T
be the α-th tangent vector on {yn = 0}, and en = (0, · · · , 0, −1)T is the
outward pointing normal to {yn ≥ 0}. Define τ α = AT eα and ν = AT en
(where A = (Dy Ψ)−1 ). By setting Gαβ = Ψ,α ·Ψ,β and G αβ = [Gαβ ]−1 , we
find that
τ α = AT eα = G αβ Ψ,β and ν = AT e n = n . (8.23)
Note that τ α and ν are also defined away from y n = 0, and τα · ν = 0 for all
α = 1, · · · , n − 1.
By (8.22), for a given function f defined in O,
∂f ∂(f ◦ Ψ) ∂(f ◦ Ψ) ∂(f ◦ Ψ)

◦ Ψ = Aji = Aαi + Ani .
∂xi ∂y j ∂y α ∂y n
Since Aαi = τiα and Ani = ni , in Φ(O) we find that
∂(f ◦ Ψ) ∂(f ◦ Ψ)
(Dx f ) ◦ Ψ = τ α +n = τ α f,α + nf,n . (8.24)
∂y α ∂y n
Some useful inequalities
The mean curvature H is defined as the trace of the second fundamental

form bαβ so that in a local chart (U, ψ),
g αβ bαβ
H ◦ψ = , bαβ = −ψ,αβ ·n .
n−1
Lemma 8.7. Suppose that (O, Ψ) is a local chart in a neighborhood of y ∈
∂Ω with coordinates (y α , y n ). If u ∈ H s (Ω) ∩ H01 (Ω) for s > 2.5, then on
∂Ω (or on {yn } = 0 in Ψ−1 (O)),
(u,ij Ni − u,ii Nj ) ◦ ψ = τjα u,αn − (n − 1)(H ◦ ψ)nj u,n ,
where H is the mean curvature of ∂Ω.
Proof. By (8.23), for y n ≥ 0, τ α = G αβ Ψ,β . Therefore, by (8.24),
u,ij ◦ Ψ = nj (ni u,nn + τiα u,αn + ni,n u,n + τi,n

α
u,α )
+ τjβ (τiα u,αβ + ni u,βn + τi,β
α
u,α + ni,β u,n ) .
Using this identity together with τ α · n = 0 and u,α = 0 if yn = 0, we find

that
h i
u,ij Ni − u,ii Nj ◦ ψ = τjβ (ni u,βn + τi,β
α
u,α + ni,β u,n )ni − τiβ (τiα u,αβ + τi,β
α
u,α + ni,β u,n )nj
= τjα u,αn − nj τiβ ni,β u,n .
The result then follows from the fact that
τiβ ni,β |yn =0 = g αβ ψ,α

i
ni,β = g αβ bαβ = (n − 1)H ◦ ψ .
Corollary 8.8. Given u ∈ H s (Ω) ∩ H01 (Ω) for s > 2.5,

Z Z ∂u 2
[u,ij Ni − u,ii Nj ]u,j dS = (n − 1) H dS .

∂Ω ∂Ω ∂N
Proof. By the definition of the surface integral,

Z N Z
X
[u,ij Ni − u,ii Nj ]u,j dS = χ` [u,ij Ni − u,ii Nj ]u,j dS
∂Ω `=1 ∂Ω
N Z
X h i p
= χ` [u,ij Ni − u,ii Nj ]u,j ◦ ψ` det(g` ) dy1 · · · dyn−1 .
`=1 ϕ(U` )
By Lemma 8.7,
h i
u,ij Ni − u,ii Nj ◦ ψ` = τjα u,αn − (n − 1)(H ◦ ψ)nj u`,n .
Therefore, by u`,i ◦ ψ` = ni u`,n on {yn = 0},
Z N Z
X p
[u,ij Ni − u,ii Nj ]u,j dS = −(n − 1) [(χ` H) ◦ ψ` ]u2`,n det(g` )dy1 · · · dyn−1 .
∂Ω `=1 ϕ(U` )
The corollary is immediately proved from this equality.
Corollary 8.9. If u ∈ H 2 (Ω) ∩ H01 (Ω), then

Z ∂u 2
2
kD uk2L2 (Ω) = k∆uk2L2 (Ω) + (n − 1) H dS . (8.25)

∂Ω ∂N
Proof. We first establish the identity is valid for all u ∈ C ∞ (Ω) ∩ H01 (Ω); a
density argument then completes the proof.
Let u ∈ C ∞ (Ω) ∩ H01 (Ω). Integrating by parts, we find that

Z Z
2 2
k∆ukL2 (Ω) = |∆u(x)| dx = u,ii (x)u,jj (x)dx
ZΩ ZΩ h i
= |D2 u(x)|2 dx − u,ij (x)Ni − u,ii (x)Nj u,j (x)dS
Ω ∂Ω
and the conclusion follows from Corollary 8.8.
Corollary 8.10. There is a constant C = C(Ω) so that

h i
kuk2H 2 (Ω) ≤ C kDuk2L2 (Ω) + k∆uk2L2 (Ω) ∀ u ∈ H 2 (Ω) ∩ H01 (Ω) . (8.26)
∂u 2
Z ∂u
Proof. We only need to estimate H dS . Since = |Du · N | ≤ |Du|,

∂Ω ∂N ∂N
Z ∂u 2
H dS ≤ C(Ω)kDuk2L2 (∂Ω) ≤ C(Ω)kDuk2H 0.25 (∂Ω) ≤ kDuk2H 0.75 (Ω) .

∂Ω ∂N
By (2.48) and Young’s inequality, we conclude that

Z ∂u 2
H dS ≤ Cδ kDuk2L2 (Ω) + δkDuk2H 1 (Ω) .

∂Ω ∂N
The identity (8.26) then follows from (8.25) and taking δ > 0 small enough.
Corollary 8.11. The norm kuk := kDukL2 (Ω) + k∆ukL2 (Ω) is an equivalent
norm in H 2 (Ω) ∩ H01 (Ω).
Remark 8.12. Similar to the proofs of Lemma 8.7 and Corollary 8.10, for
every multi-index α,
Z h i |α|+1
X
α α α
(D u),jj Ni − (D u),ij Nj (D u),i dS ≤ C kD|β| uk2L2 (∂Ω) .

∂Ω |β|=1
for all u ∈ H |α|+s (Ω) ∩ H01 (Ω). Therefore, one important conclusion of this
inequality is that the norm kuk := kukH k+1 (Ω) + k∆ukH k (Ω) is an equivalent
norm in H k+2 (Ω) ∩ H01 (Ω), or more precisely, there are constants C1 and
C2 so that for all u ∈ H k+2 (Ω) ∩ H01 (Ω),
C1 kuk2H k+2 (Ω) ≤ kuk2H k+1 (Ω) + k∆uk2H k (Ω) ≤ C2 kuk2H k+2 (Ω) . (8.27)
8.4 Vector fields with given vorticity, divergence,

and normal trace
In this section, we are concerned with obtaining a vector field u whose

vorticity curlu , divergence divu , and the normal trace u · N are given. In
other words, we look for u which satisfies
curlu = f in Ω, (8.28a)
divu = g in Ω, (8.28b)
u·N=h on ∂Ω . (8.28c)
Z
For solvability reason, we need to assume that divf = 0 and gdx =
Z Ω
hdS .
∂Ω
Let ϕ solve the elliptic equation ∆ϕ = g in Ω with Neumann boundary
∂ϕ
condition = h on ∂Ω . Then v = u − Dϕ solves
∂N
curlv = f in Ω, (8.29a)
divv = 0 in Ω, (8.29b)
v·N=0 on Ω. (8.29c)
As long as we can solve (8.29), a solution to (8.28) can be found by adding

v and Dϕ .
8.4.1 The case Ω = R3+
We first start the discussion for the case Ω = R3+ with N = (0, 0, −1) .
Existence:
By taking the curl of (8.29a), we find that −∆v = curlf in R3+ . Since
v3 = −v · N = 0 , v3 solves a Poisson equation with Dirichlet boundary
condition; hence v3 can be solved. Once v3 is obtained, by (8.29a) again, we
8.4. VECTOR FIELDS WITH GIVEN VORTICITY, DIVERGENCE, AND NORMAL TRACE349
find that
∂v1 ∂v1 ∂v3 ∂v3
=− =− − (curlv)2 = − − f2 on R2 × {x3 = 0} ,
∂N ∂x3 ∂x1 ∂x1
(8.30a)
∂v2 ∂v2 ∂v3 ∂v3
=− =− + (curlv)1 = − + f1 on R2 × {x3 = 0} .
∂N ∂x3 ∂x2 ∂x2
(8.30b)
In other words, v1 and v2 satisfy some Poisson equations with Neumann
boundary conditions (8.30); thus we can obtain v1 and v2 via solving these
Neumann problems. Note that v1 and v2 are unique up to a constant.
Even though we can solve v = (v1 , v2 , v3 ) via Dirichlet or Neumann
problems stated above, it is not clear that if v really satisfies (8.29a-b) .
However, by tangentially differentiating the Neumann boundary condition
(8.30), that is, differentiating with respect to x1 or x2 ,
∂ h ∂v1 ∂v2 i ∂ ∂v1 ∂ ∂v2
+ = +
∂x3 ∂x1 ∂x2 ∂x1 ∂x3 ∂x2 ∂x3
∂ 2 v3 ∂ 2 v3 ∂ 2 v3
= 2 + 2 + (curlf )3 = − 2 on R2 × {x3 = 0}
∂x1 ∂x2 ∂x3
which is the same as saying that
∂divv
=0 on R2 × {x3 = 0} . (8.31)
∂N
Moreover, since −∆v = curlf , ∆(divv) = 0 . In other words, divv is Har-
monic satisfying Neumann boundary condition (8.31). Therefore, divv can
only be a constant. If we insist that v ∈ H 1 (R3+ ) , then the constant must
be zero. Therefore, (8.29b) is valid.
Since divv = 0 , curlcurlv = curlf or curl(curlv − f ) = 0 in R3+ . Then
there exists a (harmonic) scalar potential ϕ such that curlv − f = Dϕ in
∂ϕ
R3+ . By (8.30b), 0 = (curlv − f )1 = on R2 × {x3 = 0} and similarly,
∂x1
∂ϕ
= 0 on R2 × {x3 = 0} . Therefore, ϕ is a constant on the boundary.
∂x2
By assuming that f ∈ L2 (R3+ ) and v ∈ H 1 (R3+ ) , ϕ is constant in R3+ ; hence
(8.29a) is valid.
Uniqueness
Suppose that both u, v ∈ H 1 (R3+ ) satisfy (8.29). Then w = u − v is ir-

rotational, divergence-free, and has zero normal trace. As a consequence,
∂ϕ
there exists a harmonic scalar potential ϕ such that w = Dϕ and = 0.
∂N
Therefore, ϕ = c for some constant c , hence u − v = 0 .
Estimates
Suppose that f ∈ H k (R3+ ) for some k ∈ N . We note that (curlf )3 ∈

H k−1 (R3+ ) . By elliptic regularity, kv3 kH k+1 (R3+ ) ≤ Ck(curlf )3 kH k−1 (R3+ ) ≤
Ckf kH k (R3+ ) ; thus
h ∂v3 i
kDv1 kH k (R3+ ) ≤ C k(curlf )1 kH k−1 (R3+ ) + f2 + ≤ Ckf kH k (R3+ ) ,

∂x1 H
k−0.5 2
(R ×{0})
h ∂v3
i
kDv2 kH k (R3+ ) ≤ C k(curlf )2 kH k−1 (R3+ ) + f1 − ≤ Ckf kH k (R3+ ) .

∂x2 H k−0.5 (R2 ×{0})

8.4.2 The case Ω is a bounded smooth domain
Now suppose that Ω is a bounded smooth domain of R3 . We first recall

from Section 8.3.1 the maps Ψ and ψ , and let Gαβ ≡ Ψ,α · Ψ,β denote the
metric tensor induced by Ψ , with G αβ denoting the inverse of Gαβ .
In a local coordinate system near the boundary, ∂j = Nj ∂N + G αβ Ψj,α ∂β ;
hence on ∂Ω ,
3
X 3
X h i
k i
curlw · N = εijk w,j N = εijk Nj w,n
k
+ g αβ ψ,α
j k
w,β Ni
i,j,k=1 i,j,k=1
3
X 2
X 2
X
j
= εijk ψ,α Ni g αβ w,β
k
= (N × ψ,α ) · g αβ w,β
i,j,k=1 α,β=1 α,β=1
2
X √
= g(δ − α)g αβ g γδ ψ,γ
k k
w,β ,
α,β,γ,δ=1
2 √
g(δ − α)g γδ ψ,γ to establish the last equality.
P
where we use N × ψ,α =
γ,δ=1
Therefore, if Ptan denotes the projection onto the tangent plane of ∂Ω , then
on ∂Ω ,
√
g(g 11 g 22 − g 12 g 12 )(ψ,2 · w,1 − ψ,1 · w,2 ) = curlw · N
√
⇔ ψ,2 · (Ptan w + wn N),1 − ψ,1 · (Ptan w + wn N),2 = gcurlw · N
√
⇔ wα,1 g2α + wα Γβ1α gβ2 − wα,2 g1α − wα Γβ2α gβ1 = gcurlw · N ;
8.4. VECTOR FIELDS WITH GIVEN VORTICITY, DIVERGENCE, AND NORMAL TRACE351
thus if DX Y denotes the covariant derivative of Y in the direction X , then
√
ψ,2 · Dψ,1 (Ptan w) − ψ,1 · Dψ,2 (Ptan w) = gcurlw · N , on ∂Ω . (8.32)
Similarly, we also have that
3 h
X 2
X 2
X i
divw = Nk wn Nk + wα Ψk,α + G αβ Ψk,α (wn Nk + wγ Ψk,γ ),β
,n
k=1 α=1 α,β,γ=1
3
X 2
X
= wn,n + G αβ Ψk,α (wγ,β Ψk,γ + wγ Ψk,βγ + Nk,β wn + Nk wn,β ) ;
k=1 α,β=1
thus if div∂Ω denotes the divergence operator on ∂Ω , then
2 h 2 i
Γβαβ wα + 2Hwn + wn,n
X X
divw = wα,α +
α=1 β=1
= div∂Ω (Ptan w) + 2Hwn + wn,n on ∂Ω . (8.33)
The case that Ω is convex
If Ω is a convex domain (which implies that H ≥ 0), we consider the following

elliptic problem
−∆w = f in Ω, (8.34a)
Ptan w = 0 on ∂Ω , (8.34b)
wn,n + 2Hwn = 0 on ∂Ω , (8.34c)
∂wn
here we recall again that wn,n = . We will look for solutions w ∈
∂N
Hτ1 (Ω) , where
n o
Hτ1 (Ω) ≡ w ∈ H 1 (Ω) Ptan w = 0 .

We note that Hτ1 (Ω) is a Banach space.

In order to solve (8.34), we try to find a weak formulation first, and this
amounts to compute
3 Z
X
i i j
− w,j ϕ N dS ∀ ϕ ∈ Hτ1 (Ω) .
i,j=1 ∂Ω
However, on ∂Ω (that is, yn = 0),

3
X 3 h
X 2
X
i
w,j Nj ϕi = j i
N wn N + wα Ψi,α
,n
i,j=1 i,j=1 α=1
2
X i
αβ
+ g Ψj,α (wn Ni + i
wσ ψ,σ ),β j i
N (ϕn N )

yn =0
α,β,σ=1
= wn,n ϕn on ∂Ω ,
we have the following
Definition 8.13. A vector-valued function w ∈ Hτ1 (Ω) is said to be a weak

solution to (8.34) if
Z Z Z
Dw · Dϕdx + 2 Hwn ϕn dS = f · ϕdx ∀ ϕ ∈ Hτ1 (Ω) . (8.35)
Ω ∂Ω Ω
Since H ≥ 0 , by (??) we know that the left-hand side of (8.35) obviously

defines a bounded, coercive bilinear form on Hτ1 (Ω) × Hτ1 (Ω) ; hence by the
Lax-Milgram theorem, there exists a unique w ∈ Hτ1 (Ω) satisfying (8.35).
If f ∈ H k (Ω) and Ω is of class H k+2 , similar techniques used to prove the
elliptic regularity imply that
kwkH k+2 (Ω) ≤ Ckf kH k (Ω) .
In this case, w solves (8.34) in the strong sense.

Now suppose that f is divergence-free. Then −∆divw = 0 in Ω . More-
over, (8.33) and (8.34b-c) imply that divw = 0 on ∂Ω . As a consequence,
divw = 0 in Ω , and this further implies that curlcurlw = f . In addi-
tion, (8.32) and (8.34b) imply that curlw · N = 0 on ∂Ω ; hence the vector
v = curlw solves (8.29) and
kvkH k+1 (Ω) ≤ Ckf kH k (Ω) .
The case that Ω is not convex
If Ω is not convex, H might be negative on some portion of ∂Ω , so there

might be no solution to (8.34) (see Exercise Problem 5.1 in Chapter 5 for
an ill-posed Robin problems). However, from the analysis above, we only
need to find a divergence-free vector w so that −∆w = f in Ω . It does
8.5. COMPENSATED COMPACTNESS AND THE DIV-CURL LEMMA353
not matter whether (8.34b) and (8.34c) hold or not. Therefore, for a given
divergence-free vector f ∈ L2 (Ω) , we can extend it so that the extension f˜
is still divergence-free and is defined on the convex domain, say B(0, R) for
some R > 0 large enough. We then consider the problem on B(0, R) instead
of Ω , or to be more precise, we consider (8.34) with B(0, R) replacing Ω .
Then the unique solution w ∈ H 2 (B(0, R)) provides us a way of constructing
a solution to (8.29), and we have
kvkH 1 (Ω) ≤ kcurlwkH 1 (Ω) ≤ Ckf kL2 (Ω) .
If f ∈ H k (Ω) with divf = 0 , the divergence-free extension f˜ in general

might not belong to H k (Ω) as well, so we cannot apply the procedure of
obtaining the elliptic regularity to show that v ∈ H k+1 (Ω) . However, we
can use the divergence and curl estimate (8.16) to obtain that
h i
kvkH k+1 (Ω) ≤ C kvkL2 (Ω) + kcurlvkH k (Ω)
h i
≤ C kwkH 1 (Ω) + k∆wkH k (Ω) ≤ Ckf kH k (Ω) .
Thus taking the scalar potential ϕ into account, we prove
Theorem 8.14. Suppose that Ω ⊆ R3 is a bounded smooth domain. Then

there exists a unique solution u ∈ H k+1 (Ω) to (8.28) if f ∈ H k (Ω) is
divergence-free,
Z g ∈Z H k (Ω) and h ∈ H k+0.5 (∂Ω) satisfies the solvability
condition gdx = hdS . Moreover, u satisfies the following estimate
Ω ∂Ω
h i
kukH k+1 (Ω) ≤ C kf kH k (Ω) + kgkH k (Ω) + khkH k+0.5 (∂Ω) .
8.5 Compensated compactness and the div-curl lemma
Lemma 8.15. Suppose vk * v and wk * w both in L2 (Ω) , and divvk and

curlwk are compact in H −1 (Ω) . Then vk · wk → v · w in D0 (Ω) .
Before proving Lemma 8.15, let us examine why this should be true.
Suppose divvk and curlwk both vanish. Then vk = curluk (v = curlu) for
some H 1 -vector field uk (u) and wk = Dpk (w = Dp) for some H 1 -scalar pk
(p). Therefore, for ϕ ∈ D(Ω) ,
Z Z
vk · wk ϕdx = curluk · Dpk ϕdx = −hcurluk , pk Dϕi ,
Ω Ω
where we use the property that divcurluk = 0 so there is not derivative

hitting curluk when integrating by parts. Now since curluk is compact in
H −1 (Ω) , and pk converges weakly in H 1 (Ω) , we find that
Z Z Z
lim vk · wk ϕdx = −hcurlu, pDϕi = curlu · Dpϕdx = v · wϕdx .
k→∞ Ω Ω Ω
We will mimic this idea to prove Lemma 8.15.
Proof. Let wk ∈ H 2 (Ω) ∩ H01 (Ω) solve
−∆wk = vk in Ω,
wk = 0 on ∂Ω .
Then vk = curlcurlwk − Ddivwk , and kwk kH 2 (Ω) ≤ Ckvk kL2 (Ω) . Moreover,
let ϕ ∈ D(Ω) . Then
−∆(ϕcurlwk ) = −curlwk ∆ϕ − 2Dϕ · Dcurlwk + ϕcurlvk .
We claim that the right-hand side, at least for a subsequence, converges

strongly in H 1 (Ω)0 . The convergence of the last term is assumed at the
very beginning, while the first term converges strongly in L2 (Ω) by the
Rellich theorem. For the second term, by the definition of the dual space
norm,
kDϕ · Dcurl(wk − w)kH 1 (Ω)0 = sup hDϕ · Dcurl(wk − w), ψiH 1 (Ω)
kψkH 1 (Ω) =1
≤ 2kDϕkW 1,∞ (Ω) kcurl(wk − w)kL2 (Ω) → 0 as k → ∞ ,
where w denotes the limit of wk . By elliptic regularity,
kϕcurl(wk − w)kH 1 (Ω) → 0 as k → ∞ .
Therefore,
Z Z Z
uk · vk ϕdx = uk · curlcurlwk ϕdx − uk · Ddivwk , ϕdx
Ω ZΩ ΩZ
= uk · curl(ϕcurlwk )dx − uk · (Dϕ × curlwk )dx

ΩZ Z Ω
+ divuk divwk ϕdx + uk · Dϕdivuk dx .
Ω Ω
Z
It is easy to see that the right-hand side converges to u · vϕdx .
Ω
8.5. COMPENSATED COMPACTNESS AND THE DIV-CURL LEMMA355
Example 8.16. Consider one dimensional Burgers equation ut + uux = 0

in R . Suppose u is the solution to the viscous Burgers equation
ut + u ux = uxx ∀x ∈ R,t > 0. (8.36)
We want to show that u converges to u in some proper sense. We consider

the two dimensional space R+ × R (treat t as another space dimension).
Then, (8.36) simply reads
(u )2
divt,x (u , ) = uxx . (8.37)
2
Multiplying (8.36) by u , we obtain
(u )3 (u )2 h (u )2 i
curlt,x (− , )= − (ux )2 . (8.38)
3 2 2 xx
Suppose that one knows that the right-hand side of (8.37) and (8.38) are
compact in H −1 , by the div-curl lemma,
(u )4 u1 u3 u22
− →− + in D0 (R) ,
12 3 4
where ui is the weak limits of (u )i . Therefore,
u4 = 4u1 u3 − 3u22 . (8.39)
Now consider the integral

Z
(u − u1 )4 dx .
R
Binomial expansion gives us
Z Z Z Z
(u − u1 ) dx = (u ) dx − 4 (u ) u1 dx + 6 (u )2 u21 dx
4 4 3
R R Z RZ R
− 4 u u31 dx + u41 dx .
R R
Passing to 0 , by (8.39) we find that

Z Z h i
4
lim (u − u1 ) φdx = u4 − 4u1 u3 + 6u2 u21 − 4u41 + u41 φdx
→0 R R Z
= −3 (u2 − u21 )2 φdx .

R
Note that the left-hand side is always non-negative, while the right-hand side
is always non-positive. Therefore, u2 = u21 or more precisely,
2
w. lim (u )2 = w. lim u .
→0 →0
8.6 Exercises
Problem 8.1. Use the calculus on the boundary to show that

Z
2
kDukL2 (Ω) + H|u · N|2 dS = kcurluk2L2 (Ω) + kdivuk2L2 (Ω)
∂Ω
n−1
X Z h i
+ bαβ (u · ψ,α )(u · ψ,β ) + 2g αβ (u · ψ,α )(u · N),β dS ,
α,β=1 ∂Ω
n−1
where H again is the mean curvature of ∂Ω , and bαβ = − g αγ g βδ ψ,γδ ·N .
P
γ,δ=1
By the identity above, deduce that
h
kuk2H 1 (Ω) ≤ C(|∂Ω|H 2.5 ) kuk2L2 (Ω) + kdivuk2L2 (Ω) + kcurluk2L2 (Ω)
i
+ ku × Nk2H 0.5 (∂Ω) ∀ u ∈ H 1 (Ω) .
Also prove that

h
i
+ k(∂u) × Nk2H k−0.5 (∂Ω) ∀ u ∈ H k+1 (Ω) .
(8.40)
where ` = max{2.5, k + 0.5} .
Problem 8.2. Prove Theorem 8.3 for the case n = 2 by the following steps:
1. Parameterize ∂Ω by (x(s), y(s)) , where s denote the arc-length, so that

the outward pointing normal is (y 0 (s), −x0 (s)) .
2. Solve (8.18). Because of (8.18b), there is a scalar function c defined

on ∂Ω so that
Dp(x(s), y(s)) = c(x(s))(x0 (s), y 0 (s)) .
Note that c can be zero at some points on ∂Ω .
3. Let B be a solution to
1
Z
−∆B = c(x)dS in Ω,
|Ω| ∂Ω
∂B
=c on ∂Ω ,
∂N
and then define v by the same formula used in the proof of Theorem
8.3. Note that now v is a scalar, not a vector.
8.6. EXERCISES 357
4. Show that vx = py and vy = −px .
5. The vector u := D⊥ v + Dp = (vy + px , −vx + py ) is an H01 (Ω)-solution

to the problem divu = f satisfying (8.17).
Problem 8.3. Let Ω ⊆ R3 be a bounded smooth domain, and u : Ω → R3 be

a vector-valued function. Suppose that u ∈ H 1 (Ω) and divu = 0 . Show that
u can be extended to the whole space so that the extension is still divergence-
free and u has compact support. In other words, find ũ ∈ H 1 (R3 ) such that
(1) divũ = 0 in R3 ; (2) ũ = u in Ω ; (3) spt(ũ) is compact.
Problem 8.4.
(1) Prove that the Leray projector P defined in Lemma ?? is linear.
(2) Prove that for u ∈ H 0 (Ω) , divP u = 0 .
Problem 8.5. Show that if u ∈ H k (Ω) is divergence free in Ω for some

k ∈ N , there exists a vector field w ∈ H k+1 (Ω) so that u = curlw and
kwkH k+1 (Ω) ≤ CkukH k (Ω) .
Chapter 9
Fluid dynamics
9.1 The equations
Let Ω ⊆ Rn be a smooth domain, occupied by a viscous incompressible fluid,

let u(x, t) denote the velocity vector field, and let p(x, t) denote the scalar
pressure function. The Navier-Stokes equations are defined as follows:
ut + (u · D)u + Dp = div(νDu) + f in Ω × (0, T ] , (9.1a)

divu = 0 in Ω × [0, T ] , (9.1b)
u=0 on ∂Ω × [0, T ] , (9.1c)
u = u0 on Ω × {t = 0} , (9.1d)
where the constant ν > 0 is the kinematic viscosity of the fluid (which, more
generally, may depend on x, t, and Du), f is the external forcing function,
and u0 is the initial velocity. The nonlinearity (u · D)u is the advection term
defined by
n
X ∂u
(u · D)u = uj .
∂xj
j=1
Note that (u · D)u is, in fact, the directional derivative of u in the direction
u , and is hence often expressed as Du u . Condition (9.1b) is often called
the divergence-free or the incompressibility condition, while the boundary
condition (9.1c) is known as the no-slip condition.
When the viscosity ν is large and diffusion dominates advection, the
358
9.2. INCOMPRESSIBLE EULER EQUATIONS ON A FIXED DOMAIN359
nonlinearity if ignored and (9.1) reduces to the linear Stokes equations:
ut − ν∆u + Dp = f in Ω × (0, T ] , (9.2a)

divu = 0 in Ω × [0, T ] , (9.2b)
u=0 on ∂Ω × [0, T ] , (9.2c)
u = u0 on Ω × {t = 0} . (9.2d)
On the other hand, when the viscosity ν is very small, and advection
dominates diffusion, we obtain the equations of a perfect or ideal incom-
pressible fluid, given by the Euler equations:
ut + (u · D)u + Dp = f in Ω × (0, T ] , (9.3a)

divu = 0 in Ω × [0, T ] , (9.3b)
u·N=0 on ∂Ω × [0, T ] , (9.3c)
u = u0 on Ω × {t = 0} . (9.3d)
9.2 Incompressible Euler equations on a fixed do-

main
9.2.1 Flow on a bounded and fixed domain
Let Ω denote a bounded, smooth, open subset of Rn with n = 2, 3, and let N

denote the outward-pointing unit normal to ∂Ω. We consider the following
system of equations describing the flow of an inviscid incompressible fluid:
ut + (u · D)u + Dp = 0 in Ω × (0, T ] , (9.4a)

divu = 0 in Ω × [0, T ] , (9.4b)
u·N=0 on ∂Ω × [0, T ] , (9.4c)
u = u0 on Ω × {t = 0} , (9.4d)
where the vector field u = (u1 , ..., un ) denotes the velocity of the fluid and
the scalar function p denotes the pressure in the fluid. The equation (9.4a)
denotes conservation of momentum, (9.4b) states that the fluid is incom-
pressible so that the fluid flow preserves volume, the boundary condition
(9.4c) states that fluid cannot leave the fluid container Ω, and (9.4d) pre-
scribes the initial velocity field u0 , which must also be divergence-free. The
ith component of the advection term can be expressed as [(u·D)u]i = ui ,j uj .
360 CHAPTER 9. Fluid dynamics
By first taking the divergence of (9.4a) and second computing its inner-
product with N on ∂Ω, we obtain the elliptic equation
−∆p = ui ,j uj ,i in Ω × [0, T ] , (9.5a)

∂p
= −ui ,j uj Ni on ∂Ω × [0, T ] . (9.5b)
∂N
An important observation is that the right-hand side of (9.5b) is, in fact,

quadratic in u. To see this, we rely on the identity (8.24), and write (in a
local chart (U, ψ) with n = N ◦ ψ)
−ui ,j ◦ ψ = −ui ,α g αβ ψ j ,β −ui ,n nj .
With τ α denoting g αβ ψ,β , it follows that
−(ui ,j uj Ni ) ◦ ψ = −(ui ,α τjα + ui ,n nj ) ni uj

= −ui ,α ni uj τjα = ui ni ,α (u · τ α )
= (uk τkβ ψ i ,β ) ni ,α (u · τ α )
= (u · τkβ )ψ i ,β g στ bασ ψ i ,τ (u · τ α ) = bαβ (u · τ α ) (u · τ β ) ;
thus (9.5b) can be written as
∂p
= b(u · τ, u · τ ) = b(u, u) . (9.5b’)
∂N
Standard elliptic estimates for the Neumann problem show that for s ≥ 0,
h i
kpkH s+1 (Ω) ≤ C kui ,j uj ,i kH s−1 (Ω) + kb(u, u)k 1 . (9.6)
H s− 2 (∂Ω)
9.2.2 The space of divergence-free vector fields on Ω
Lemma 9.1 (Normal trace lemma). Suppose that u ∈ L2 (Ω) and divu ∈
1
L2 (Ω). Then u · N ∈ H − 2 (∂Ω) and for a generic constant C > 0 depending
on Ω,
h i
ku · Nk 1
−2 ≤ C kukL2 (Ω) + kdivukL2 (Ω) .
H (∂Ω)
Proof. By density, assume that u ∈ C ∞ (Ω) and let φ be an arbitrary func-

tion in H 1 (Ω). According to Theorems 2.153 and 2.156, the trace operator
§9.2 Incompressible Euler equations 361
1
T : H 1 (Ω) → H 2 (∂Ω) is a continuous surjection so we may identify φ with
1
an H 2 (∂Ω) function. It follows that

ku · Nk 1
−2 = sup hu · N, φi ,
H (∂Ω) kφk =1
1
H 2 (∂Ω)
1 1
where h·, ·i denote the duality pairing between H 2 (∂Ω) and H − 2 (∂Ω). By
the divergence theorem,
Z

hu · N, φi = u · N φ dS

∂Ω
Z
= divu φ + u · Dφ dx ≤ kdivukL2 (Ω) + kukL2 (Ω) kφkH 1 (Ω) .

Ω
Thanks to Lemma 9.1 we can make the following definition of an impor-

tant subspace of H s (Ω):
Definition 9.2 (Space of divergence-free vector fields with vanishing normal

trace). For k ≥ 0, set
V k = u ∈ H k (Ω) divu = 0 in Ω , u · N = 0 on ∂Ω ,

where divu is taken in the distributional sense. The space V k is a Hilbert

space when endowed with the H k (Ω)-norm.
The Hodge decomposition shows that when s ≥ 0,
H s (Ω) = V s ⊕L2 D[H s+1 (Ω)] ,
and there are natural projections onto the right-hand sides of this decom-
posion.
Definition 9.3 (Leray Projection). We define the projection operator P :

H s (Ω) → V s by P(w) = w −Dp, where p is the unique solution of the elliptic
equation
∂p
∆p = divw in Ω and = w · N on ∂Ω .
∂N
The divergence theorem shows that the solvability condition for this Neu-
mann problem is always satisfied.
The Leray projection allows us to reexpress the Euler equations as the

following system:

ut + P Du u = 0 in Ω × (0, T ] , (9.7a)
u = u0 on Ω × {t = 0} . (9.7b)
9.2.3 The Lagrangian flow of a velocity field
The passage to the so-called Lagrangian coordinates is extremely important

in the study of hyperbolic PDE.
Definition 9.4 (Lagrangian coordinates). We set η(x, t) to be the solution

of the ODE
∂t η(x, t) = u(η(x, t), t) for 0 < t ≤ T , and η = e on {t = 0} , (9.8)
where e(x) = x denotes the identity map on Ω. Given a particle labeled x at

t = 0, η(x, t) gives the position of this particle at time t ∈ (0, T ]. Whenever,
the velocity field u is Lipschitz continuous in x, uniformly in t, there exists
a unique solution η. The time T > 0 depends on u.
Lemma 9.5. If u · N = 0 on ∂Ω, then η(t)(∂Ω) ⊆ ∂Ω for t ∈ [0, T ].
Proof. Let x be any arbitrary and fixed element of ∂Ω. We must prove
that η(x, t) ∈ ∂Ω for all t ∈ [0, T ]. It suffices to show that each vector
ηt (x, t) is in the tangent space Tη(x,t) ∂Ω. This follows from the fact that
ηt (x, t) · N(η(x, t)) = [u · N](η(x, t)) = 0.
Lemma 9.6. If divu = 0 in Ω × [0, T ], then det Dη = 1 in Ω × [0, T ].
Proof. Let A(x, t) = [Dη(x, t)]−1 denote the n × n inverse matrix, let J =
det Dη, and let
a(x, t) = J(x, t)A(x, t)
denote the matrix of cofactors. We compute the time-derivative of J as
Jt = asr ηtr ,s = JAsr ηtr ,s = J divu ◦ η ,
the last equality following from the chain rule. Since J(x, 0) = 1, given that
η(x, 0) = x, we see that J(x, t) = 1 whenever divu = 0.
Proposition 9.7 (A polynomial-type inequality). For a constant M0 ≥ 0,

suppose that f (t) ≥ 0, t 7→ f (t) is continuous, and
f (t) ≤ M0 + C t P(f (t)) , (9.9)
where P denotes a polynomial function, and C is a generic constant. Then

for t taken sufficiently small (independently of the function f ≥ 0 satisfying
(9.9)), we have the bound
f (t) ≤ 2M0 .
We will see many applications of Proposition 9.7, which can be viewed

as a generalization of the nonlinear Gronwall inequality. In order to see the
first use of this result, it is convenient to first make the following
n
Definition 9.8 (The group of diffeomorphisms of Ω). For s > + 1, let
2
Ds (Ω) = ψ ∈ H s (Ω; Ω) ψ −1 exists, and ψ −1 ∈ H s (Ω; Ω) .

Ds (Ω) is the group of volume-preseving diffeomorphisms with group multi-

plication given by composition of maps on the right.
For each η ∈ Ds (Ω), the tangent space at η of this group is given by
Tη Ds (Ω) = v := u ◦ η u ∈ H s (Ω) and u · N = 0 on ∂Ω .

The tangent bundle T Ds (Ω) = ∪η∈Ds (Ω) Tη Ds (Ω) and is thus locally isomor-
phic to the space of pairs of “positions” and “velocities” (η, v).
Definition 9.9 (Subgroup of volume-preserving diffeomorphisms of Ω). For

n
s > + 1, we set
2
s
Dvol (Ω) = {η ∈ Ds (Ω) | det Dη = 1} .
s (Ω), the tangent space at η of this subgroup is given by

For each η ∈ Dvol
s
(Ω) = v := u ◦ η u ∈ V s .

Tη Dvol
Lemma 9.10. Suppose that u ∈ C([0, T1 ]; H 3 (Ω)) with u · N = 0 on ∂Ω ×

[0, T1 ]. Then there exists a unique solution η(x, t) of (9.8) with η ∈ C 1 ([0, T ]; Ds (Ω))
for some T ≤ T1 , and max kη(t)k2H 3 (Ω) ≤ 2kek2H 3 (Ω) .
[0,T ]
Proof. By the Sobolev embedding theorem, u is uniformly Lipschitz contin-

uous, so that there exists a unique solution η to (9.8). By the fundamental
theorem of calculus (with η(t) denoting η(·, t)),
Z t
η(t) = e + u ◦ η(s)ds ,
0
so that
X Z t
max kη(t)k2H 3 (Ω) ≤ kek2H 3 (Ω) + max kDα (u ◦ η)k2L2 (Ω) ds
[0,T ] [0,T ]
|α|≤3 0
≤ kek2H 3 (Ω) + T 2 max ku(t)k2H 3 (Ω) P max kη(t)k2H 3 (Ω) ,

[0,T ] [0,T ]
where we have used the chain rule, the Sobolev embedding theorem, and
Hölder’s and Jensen’s inequality to obtain the second inequality. Proposition
9.7 then provides the desired result for T > 0 taken small enough.
n
Remark 9.11. If for s > + 1, u ∈ C([0, T1 ]; H s (Ω) with u · N = 0 on
2
∂Ω × [0, T1 ], then a generalization of the argument given in Lemma 9.10
shows that the flow η ∈ C 1 ([0, T ; Ds (Ω)) for T ≤ T1 . If u ∈ C([0, T1 ]; V s ),
then according to Lemma 9.6, each η ∈ C 1 ([0, T ]; Dvol
s (Ω)).
Remark 9.12. Since the flow map η(t) starts at the identity, for T >
0 taken sufficiently small, η(t) is a near-identity transformation, so that
max[0,T ] kη(t) − ek23 ≤ δ(T ) where δ(T ) → 0 as T → 0. Thus, by the inverse
function theorem η −1 (t) exists. Moreover, since Dη −1 ◦ η = [Dη]−1 , and
since [Dη]−1 = cof[Dη]/ det Dη, we see that each component of [Dη]−1 is a
n
polynomial function of partial derivatives of η. It follows that for s > + 1,
2
the map η 7→ η −1 : Ds → Ds is continuous.
Lemma 9.13. Let η ∈ D3 (Ω) and f ∈ H 3 (Ω). If kηkH 3 (Ω) ≤ C, then

kf ◦ ηkH 3 (Ω) ≤ Ckf kH 3 (Ω) .
Proof. The chain rule shows that
D(f ◦ η) = Df ◦ η Dη ,
D2 (f ◦ η) = D2 f ◦ η Dη Dη + Df ◦ η D2 η ,
D3 (f ◦ η) = D3 f ◦ η Dη Dη Dη + 3D2 f ◦ η D2 η Dη + Df ◦ η D3 η .
We then see that

Z
3
kD (f ◦ η)k2L2 (Ω) ≤ |D3 f ◦ η|2 |Dη(x)|6 dx
Ω Z
+ 3 |D2 f ◦ η|2 |D2 η(x)|2 |Dη(x)|2 dx
Ω Z
+ |Df ◦ η|2 |D3 η(x)|2 dx
Z Ω
6
≤ kDηkL∞ (Ω) |D3 f |2 | det Dη(x)|dx
Ω Z
2
+ 3kDf kL∞ (Ω) |D3 η(x)|2 dx
Ω
+ 3kDηk2L∞ (Ω) kD2 ηk2L6 (Ω) kD2 f ◦ ηk2L3 (Ω)
≤ P (kηkH 3 (Ω) )kf kH 3 (Ω) .
The lower-order terms are estimated in a similar fashion.
9.2.4 Existence and uniqueness of solutions of the incom-

pressible Euler equations on fixed and bounded domain
We begin with the case that initial data has H 3 -class regularity.
Theorem 9.14. For u0 ∈ V 3 , there exists a unique solution u to (9.4) such

that
u ∈ C([0, T ]; V 3 ) ∩ C 1 ([0, T ]; V 2 ) ,
where T > 0 depends on ku0 kH 3 (Ω) .
Proof. Consider the subset of the Banach space C([0, T ], V 3 ) defined by
CT = {v ∈ C([0, T ]; H 3 (Ω) | v · N = 0 on ∂Ω × [0, T ] ,

max kv(t)k2H 3 (Ω) ≤ ku0 k2H 3 (Ω) + 1} ,
t∈[0,T ]
with T and M to be subsequently fixed. Choose v̄ ∈ CT , and set η̄ = e +

Rt 0 0 2
0 v̄(t )dt . The fundamental theorem of calculus shows that max[0,T ] kη̄(t)k3 ≤
C for T > 0 taken sufficiently small, and hence η̄(t) ∈ D3 (Ω) for each
t ∈ [0, T ] by the inverse function theorem.
Consider the linear problem
∂t ui + ui ,j ūj + pu ,i = 0 in Ω × (0, T ] , (9.10a)

divu = 0 in Ω × [0, T ] , (9.10b)
u·N=0 on ∂Ω × [0, T ] , (9.10c)
u = u0 on Ω × {t = 0} , (9.10d)
where the pressure function pu satisfies the linear elliptic equation
−∆pu = ui ,j ūj ,i in Ω × [0, T ] , (9.11a)

∂pu
= b(u, ū) on ∂Ω × [0, T ] , (9.11b)
∂N
and verifies the estimate
kpu kH 4 (Ω) ≤ C kui ,j ūj ,i kH 2 (Ω) + kb(u, ū)kH 2.5 (∂Ω)

≤ CkukH 3 (Ω) kūkH 3 (Ω) . (9.12)
If we define the Lagrangian velocity v associated to the solution u of (9.10)

by v := u ◦ η̄, then (9.10a) can be written as
vt + Dpu (η̄) = 0 . (9.10a’)
so that by the fundamental theorem of calculus,

Z t
v(t) = u0 − Dpu (η̄)ds . (9.10a”)
0
For v̄ ∈ CT fixed, we next construct a solution to (9.10a’) as the fixed-

point of the map v 7→ ṽ , where v ∈ CT is given and
Z t
ṽ(t) = u0 − Dpu (η̄)ds .
0
The chain rule, the Sobolev embedding theorem, and Hölder’s and Jensen’s
inequality show that
max kṽ(t)k2H 3 (Ω) ≤ ku0 k2H 3 (Ω) + T max kpu k2H 4 (Ω) P ( max kη̄(t)k2H 3 (Ω) )
t∈[0,T ] t∈[0,T ] t∈[0,T ]
According to Lemma 9.10, maxt∈[0,T ] kη̄(t)k2H 3 (Ω) ≤ C, so that with (9.12),
max kṽ(t)k2H 3 (Ω) ≤ ku0 k2H 3 (Ω) + C T max ku(t)k2H 3 (Ω)

t∈[0,T ] t∈[0,T ]
≤ ku0 k2H 3 (Ω) + C T max kv(t)k2H 3 (Ω) ≤ ku0 k2H 3 (Ω) + C T ,

t∈[0,T ]
so that for T > 0 sufficiently small, ṽ ∈ CT . Next, let v (a) ∈ CT for a = 1, 2,

(a)
define u(a) := v (a) ◦ η̄ −1 , and let pu denote the solution to
−∆p(a) (a) i j
u = [u ] ,j ū ,i in Ω × [0, T ] ,
(a)
∂pu
= b(u(a) , ū) on ∂Ω × [0, T ] ,
∂N
and Z t
(a)
ṽ (t) = u0 − Dp(a)
u (η̄)ds .
0
Then, following our estimates above, we see that
max kṽ (1) (t) − ṽ (2) (t)k ≤ C T max kv (1) (t) − v (2) (t)k .
t∈[0,T ] t∈[0,T ]
The contraction mapping principle provides a unique fixed-point to the map

v 7→ ṽ in CT which is a solution to (9.10a”), and hence (9.10a) and (9.10d)
are satisfied. The conditions (9.10b) and (9.10c) are then verified by the
definition of pu . Having obtained a solution u ∈ V 3 to the linear problem
(9.10), we can now construct a solution to the incompressible Euler equations
(9.4) as the fixed-point of the map v̄ 7→ v in the space CT for T > 0 taken
sufficiently small.
From (9.10a”), we once again see that
max kv(t)k2H 3 (Ω) ≤ ku0 k2H 3 (Ω) + C T max kv(t)k2H 3 (Ω) ,

t∈[0,T ] t∈[0,T ]
so that v ∈ CT for T > 0 small enough.

Next, suppose that v̄1 and v̄2 are both in CT , and for a = 1, 2, set
Rt
η̄a (t) = e + 0 v̄a , ūa = v̄a ◦ η̄a−1 , ua = va ◦ η̄a−1 ,and let Pa := paua solve
−∆paua = uia ,j ūja ,i in Ω × [0, T ] , (9.14a)

∂paua
= b(ua , ūa ) on ∂Ω × [0, T ] , (9.14b)
∂N
It follows that with
Z t Z t
v1 (t) = u0 − DP1 (η̄1 )ds and v2 (t) = u0 − DP2 (η̄2 )ds ,
0 0
we have the estimate
max kv1 (t) − v2 (t)k2H 3 (Ω) ≤ T 2 max kDP1 ◦ η̄1 − DP2 ◦ η̄2 k2H 3 (Ω) .
t∈[0,T ] t∈[0,T ]
It remains to show that

kDP1 ◦ η̄1 − DP2 ◦ η̄2 kH 3 (Ω) ≤ C kv1 − v2 kH 3 (Ω) + kv̄1 − v̄2 kH 3 (Ω) , (9.15)
as the contraction mapping principle would then conclude the proof.

Setting ξ = η̄1 ◦ η̄2−1 , Lemma 9.13 shows that
kDP1 ◦ η̄1 − DP2 ◦ η̄2 kH 3 (Ω) ≤ CkDP1 ◦ ξ − DP2 kH 3 (Ω) .
In order to estimate this right-hand side, we will rely on the elliptic estimate
(8.16) and show that

kdiv(DP1 ◦ ξ − DP2 )kH 2 (Ω) ≤ C kv1 − v2 kH 3 (Ω) + kv̄1 − v̄2 kH 3 (Ω) ,
(9.16)

k curl(DP1 ◦ ξ − DP2 )kH 2 (Ω) ≤ C kv1 − v2 kH 3 (Ω) + kv̄1 − v̄2 kH 3 (Ω) ,
(9.17)

k(DP1 ◦ ξ − DP2 ) · NkH 2.5 (∂Ω) ≤ C kv1 − v2 kH 3 (Ω) + kv̄1 − v̄2 kH 3 (Ω) ,
(9.18)
and

kDP1 ◦ ξ − DP2 kL2 (Ω) ≤ C kv1 − v2 kH 3 (Ω) + kv̄1 − v̄2 kH 3 (Ω) . (9.19)
For this purpose, we write
div(DP1 ◦ ξ) = ∆P1 ◦ ξ + P1 ,ij ◦ξ (ξ j ,i −δij ) ,
and we rewrite (9.14) as

h i
−∆Pa = vai ,r [Āa ]rj v̄aj ,s [Āa ]si ◦ η̄a−1 in Ω × [0, T ] , (9.20a)
∂Pa
= b(va , v̄a ) ◦ η̄a−1 on ∂Ω × [0, T ] . (9.20b)
∂N
Hence,
kdiv(DP1 ◦ ξ − DP2 )kH 2 (Ω)

≤ k∆P1 ◦ ξ − ∆P2 kH 2 (Ω) + kP1 ,ij ◦ξ(ξ j ,i −δij )kH 2 (Ω)

≤ C k∆P1 ◦ η̄1 − ∆P2 ◦ η̄2 kH 2 (Ω) + kP1 ,ij ◦ξkH 2 (Ω) k(η̄1j ,i −η̄2j ,i )kH 2 (Ω)

≤ C kv1i ,r [Ā1 ]rj v̄1j ,s [Ā1 ]si − v2i ,r [Ā2 ]rj v̄2j ,s [Ā2 ]si kH 2 (Ω) + k(η̄1j ,i −η̄2j ,i )kH 2 (Ω) ,
9.3. THE STOKES EQUATIONS 369
the last inequality following from (9.20a). It follows that (9.16) holds; sim-
ilarly, (9.18) holds by replacing the divergence operator with the normal
trace, and using (9.20b). The inequality (9.17) follows from the fact that
curl DP1 = 0.
In order to obtain the inequality (9.19), we see that
kDP1 ◦ ξ − DP2 kL2 (Ω) ≤ CkDP1 ◦ η̄1 − DP2 ◦ η̄2 kL2 (Ω) .
With Qa = Pa ◦ η̄a for a = 1, 2, we write (9.20) as

−[Aa ]ji [Aa ]ki Qa ,k ,j = vai ,r [Āa ]rj v̄aj ,s [Āa ]si in Ω × [0, T ] , (9.21a)
Qa ,k [Aa ]ki [Aa ]ji Nj = b(va , v̄a ) on ∂Ω × [0, T ] . (9.21b)
Since P1 ,i ◦η̄1 −P2 ,i ◦η̄2 = Q1 ,k [A1 ]ki −Q2 ,k [A2 ]ki , elliptic estimates for (9.21)
show that (9.19) holds.
An induction argument show that the more general statement of the

theorem holds:
Theorem 9.15 (Existence and uniqueness for the incompressible Euler

n
equations). For u0 ∈ V s , s > + 1, there exists a unique solution u to
2
(9.4) such that
u ∈ C([0, T ]; V s ) ∩ C 1 ([0, T ]; V s−1 ) ,
where T > 0 depends on ku0 kH s (Ω) .
9.3 The Stokes equations

9.3.1 The steady state case
In this section, we are concerned with the steady state Stokes equations
(with ν = 1)
−∆u + Dp = f in Ω × (0, T ) , (9.22a)

divu = 0 in Ω × (0, T ) , (9.22b)
u=0 on ∂Ω × (0, T ) . (9.22c)
Note that since u and p are time independent, we need the initial condition
(9.2d) no more.
Similar to V k defined in the previous section, for k ∈ N we define the

space W k by
n o
W k ≡ u ∈ H01 (Ω; Rn ) ∩ H k (Ω; Rn ) divu = 0

with the usual H k (Ω)-norm. W k is a Banach space for all k ∈ N.

Suppose that ϕ ∈ W 1 , then (Dp, ϕ)L2 (Ω) = 0 if p for p ∈ H 1 (Ω). This
observation leads to the following definition.
Definition 9.16. A vector valued function u ∈ W 1 is said to be a weak

solution to (9.22) if
(Du, Dϕ)L2 (Ω;Rn ) = hf, ϕiW 1 ∀ ϕ ∈ W 1. (9.23)
It is easy to see from the Lax-Milgram Theorem that a unique weak

solution to (9.22) exists. Moreover, the weak solution u satisfies
kukH01 (Ω) = kDukL2 (Ω) ≤ kf kH −1 (Ω) .
The question of how to recover the pressure function is given by the following
Lemma 9.17 (Lagrange multiplier lemma). Suppose that T : H01 (Ω; Rn ) →

R is a bounded linear functional such that T ϕ = 0 for all ϕ ∈ W 1 . Then
there exists p ∈ L2 (Ω)/R, unique modulo a constant, such that
T ϕ = (p, divϕ)L2 (Ω) ∀ ϕ ∈ H01 (Ω; Rn ).
Moreover,
kpkL2 (Ω) ≤ kT kB(H01 (Ω;Rn ),R) . (9.24)
Note, that we are using k · kB(H01 (Ω;Rn ),R) to denote the operator norm.
Proof. Given p ∈ L2 (Ω)/R, define a linear functional Tp : H01 (Ω; Rn ) → R

by Tp ϕ = (p, divϕ)L2 (Ω) . Then
kTp kB(H01 (Ω;Rn ),R) = sup |(p, divϕ)L2 (Ω) | ≤ kpkL2 (Ω) ;
kDϕkL2 (Ω) =1
thus Tp is bounded. By the Riesz representation theorem, there exists Qp ∈

H01 (Ω) such that
Tp ϕ = (p, divϕ)L2 (Ω) = (Qp, ϕ)H01 (Ω) ,

§9.3 The Stokes equations 371
and kQpkH01 (Ω) = kTp kB(H01 (Ω;Rn ),R) ≤ CkpkL2 (Ω) . Therefore, Q : L2 (Ω)/R →
H01 (Ω) is a bounded linear functional.
On the other hand, for this given p ∈ L2 (Ω)/R, by Theorem 8.3 there
exists ϕ ∈ H01 (Ω) such that divϕ = p. Therefore,
Tp ϕ = kpk2L2 (Ω) ≤ kQpkH01 (Ω) kpkH01 (Ω)
which implies that
kQpkH01 (Ω) ≥ kpkL2 (Ω) . (9.25)
As a consequence, R(Q), the range of Q, is closed (see the proof of the

Lax-Milgram Theorem for the detail). Since R(Q)⊥ = W 1 ,
H01 (Ω) = R(Q) ⊕H01 (Ω) R(Q)⊥ = R(Q) ⊕H01 (Ω) W 1 .
Since T : H01 (Ω; Rn ) → R is bounded, by the Riesz representation theorem

there exists ψ ∈ H01 (Ω; Rn ) such that T ϕ = (ψ, ϕ)H01 (Ω) and kψkH01 (Ω;Rn ) =
kT kB(H01 (Ω;Rn ),R) . By the decomposition above, ψ = ψ1 + ψ2 for some ψ1 ∈
R(Q) and ψ2 ∈ W 1 . Let ψ1 = Qp. For ϕ ∈ H01 (Ω), write ϕ = ϕ1 + ϕ2 for
some ϕ1 ∈ R(Q) and ϕ2 ∈ W 1 . Then
T ϕ = T ϕ1 = (ψ, ϕ1 )H01 (Ω;Rn ) = (ψ1 , ϕ1 )H01 (Ω;Rn ) = (ψ1 , ϕ)H01 (Ω;Rn )

= (Qp, ϕ)H01 (Ω;Rn ) = (p, divϕ)L2 (Ω) .
Moreover, (9.25) implies that
kpkL2 (Ω) ≤ kQpkH01 (Ω) = kψ1 kH01 (Ω) ≤ kψkH01 (Ω) = kT kB(H01 (Ω;Rn ),R) .
Given f ∈ H −1 (Ω; Rn ), let u be the weak solution to (9.22). Define

T : H01 (Ω; Rn ) → R by T ϕ = (Du, Dϕ)L2 (Ω) − hf, ϕiH01 (Ω;Rn ) . Then
kT kB(H01 (Ω;Rn ),R) = sup |T ϕ| ≤ kDukL2 (Ω) + kf kH −1 (Ω) ≤ 2kf kH −1 (Ω)

kDϕkL2 (Ω) =1
implies that T : H01 (Ω; Rn ) → R is a bounded linear functional. Moreover,

T ϕ = 0 for all ϕ ∈ W 1 . Therefore, Lemma 9.17 provides the existence of a
pressure p ∈ L2 (Ω) such that
(Du, Dϕ)L2 (Ω) − hf, ϕiH01 (Ω;Rn ) = (p, divϕ)L2 (Ω) ∀ ϕ ∈ H01 (Ω; Rn ),
and this p satisfies
kpkL2 (Ω) ≤ kT kB(H01 (Ω;Rn ),R) ≤ 2kf kH −1 (Ω) . (9.26)
Based on the discussion above, we have the following alternative definition

of weak solution.
Definition 9.18. The pair (u, p) ∈ W 1 ×L2 (Ω) is said to be a weak solution
to (9.22) if
(Du, Dϕ)L2 (Ω) + (p, divϕ)L2 (Ω) = hf, ϕiH01 (Ω;Rn ) ∀ ϕ ∈ H01 (Ω; Rn ). (9.27)
Regularity theory for the Stokes problem
Now suppose that f ∈ L2 (Ω). Similar to the regularity theory of Section

5.9, we expect that inversion of the Stokes operator provides a two-derivative
gain of regularity with respect to the forcing function f . In particular, for
f ∈ L2 (Ω), we should be able to show that u ∈ W 2 .
Indeed, this is the case, and the proof that u ∈ W 2 is somewhat similar
to the proof of elliptic regularity in Section 5.9, once we agree to use the
variational formulation (9.27) rather then (9.23). The primary advantage
of the form (9.27) comes from the removal of the divergence-free constraint
on the test functions. Notice that if we set v = η ∗ ζ 2 D2 (η ∗ u) as we

introduced in Section 5.9 for the proof of interior regularity, then the test
function v ∈ H01 (Ω; Rn ) but the presence of the cut-off function does not
preserve the divergence-free constraint and divv 6= 0.
Theorem 9.19. Suppose that ∂Ω is of class C m+1 for some m ∈ N. Then

for any f ∈ H m−1 (Ω), there exists a unique solution (u, p) ∈ W m+1 ×
H m (Ω)/R to the steady Stokes problem
−∆u + Dp = f in Ω × (0, T ) ,
divu = 0 in Ω × (0, T ) ,
u=0 on ∂Ω × (0, T ) .
Moreover, (u, p) satisfies
kukH m+1 (Ω) + kpkH m (Ω) ≤ Ckf kH m−1 (Ω) . (9.29)

Proof. We first prove that u ∈ W 2 ; the general case follows by an induction

argument.
For the proof, we will use the same cut-off functions ζ, ζm , and local
charts θm as we introduced in Section 5.9.
Step 1. Interior estimates.
Let v = η ∗ ζ 2 (η ∗ u),k ,k . The use of v as a test function in (9.27)

implies that
− (Du, Dv)L2 (Ω) = −(f, v)L2 (Ω) + (p, divv)L2 (Ω)

≤ Cδ kf k2L2 (Ω) + δkζD2 (η ∗ u)k2L2 (Ω) − (p, divv)L2 (Ω) . (9.30)
Since divu = 0,
Z
p η ∗ [ζ 2 (η ∗ ui ),k ],k ,i dx

(p, divv)L2 (Ω) =
Ω
Z
p η ∗ [ζζ,i (η ∗ ui ),k ],k dx;

=2 (9.31)
Ω
thus by (9.26) and Young’s inequality,

h i
(p, divv)L2 (Ω) ≤ CkpkL2 (Ω) kukH 1 (Ω) + kζD2 (η ∗ u)kL2 (Ω)
≤ Cδ kf k2L2 (Ω) + δkζD2 (η ∗ u)k2L2 (Ω) . (9.32)
On the other hand, just as in Section 5.9, we have that
−(Du,Dv)L2 (Ω) = D(η ∗ u),k , D[(ζ 2 (η ∗ u),k ] L2 (Ω)

≥ kζD2 (η ∗ u)k2L2 (Ω) − CkD(η ∗ u)kL2 (Ω) kζD2 (η ∗ u)k2L2 (Ω)
≥ (1 − δ)kζD2 (η ∗ u)k2L2 (Ω) − Cδ kuk2H 1 (Ω)
≥ (1 − δ)kζD2 (η ∗ u)k2L2 (Ω) − Cδ kf k2L2 (Ω) .
Therefore, by choosing δ > 0 small enough, we conclude that
kζD2 (η ∗ u)k2L2 (Ω) ≤ Ckf kL2 (Ω)
for some constant C = C(kDζkL∞ (Ω) ). Since the bound on the right-hand
side is independent of , we may pass to a weak limit (of a subsequence) and
obtain that
kζD2 uk2L2 (Ω) ≤ Ckf kL2 (Ω) . (9.33)

2 (Ω).
We note that estimate (9.33) implies that u ∈ Hloc
Given ψ ∈ H 2 (Ω), let ϕ = ζDψ in (9.27). With Defu denoting the
1
symmetric part of Du; that is, Defu = Du + (Du)T , we find that

2

(ζp, ∆ψ)L2 (Ω) = (ζf − pDζ, Dψ)L2 (Ω) − Du, D(ζDψ) L2 (Ω)
= (ζf − pDζ − DefuDζ, Dψ)L2 (Ω) . (9.34)
Identity (9.34) is the same as saying that q = ζp is a distributional solution

to the elliptic equation
−∆q = div(ζf − pDζ − DefuDζ) in Ω, (9.35a)

q=0 on ∂Ω . (9.35b)
However, since div(ζf − pDζ − DefuDζ) ∈ H −1 (Ω), there exists a weak

solution q̃ ∈ H01 (Ω) to (9.35). Moreover, since (q − q̃, ∆ψ)L2 (Ω) = 0 for all
ψ ∈ C02 (Ω), q̃ must agree with q. Therefore, ζp ∈ H01 (Ω) satisfying
kζpkH 1 (Ω) ≤ Ckdiv(ζf − pDζ − DefuDζ)kH −1 (Ω) ≤ Ckf kL2 (Ω) (9.36)
for some constant C = C(kDζkL∞ (Ω) ), and estimate (9.36) implies that
1 (Ω).
p ∈ Hloc
2 (Ω) and p ∈ H 1 (Ω), we can integrate by
Having established u ∈ Hloc loc
parts (to move the derivative on the test function back onto u and p in
(9.23)) and obtain that (f + ∆u − Dp, ϕ)L2 (Ω) = 0 for all ϕ ∈ C01 (Ω). This
implies that
−∆u + Dp = f a.e. in Ω. (9.37)
Step 2. Boundary estimates for u ◦ θm .

Define A = (Dθm )−1 , ξm = ζm ◦ θm and um = u ◦ θm as in the proof of
−1
the elliptic regularity. Let v i = θm,`
i 2 Λ (A` umj ),
Λ ξm j α ,α ◦ θm be a test
function in (9.27). Using the commutator notation,

A, B u = A B u − B A u ,
we have that
(v ◦ θm )i = θm,`
i 2
Λ (A`j umj ),α ,α

Λ ξm
i 2
Λ (A`j ,α umj ) ,α +2θm,`
i
Λ ξm ξm ,α Λ (A`j umj ,α )

= θm,` Λ ξm
i 2
∂α Λ , A`j umj ,α + θm,ì 2 `
Aj ,α Λ umj ,α )

+ θm,` Λ ξm Λ (ξm
i 2 `
Aj (Λ umj ),αα .

+ θm,` Λ ξm
Applying the commutation estimate (5.46) to estimate the third term on

the right-hand side, we find that
h ` 2 i
mj

kvkL2 (Ω) ≤ C kukH 1 (Ω) + Aj ξm (Λ u
,αα L2 (B+ (0,rm ))

h i
¯
≤ C kf kL2 (Ω) + kξm ∂DΛ um
k L (B+ (0,rm )) ,
2

where C = C kDζm kL∞ (B+ (0,rm )) , kDθm kL∞ (B+ (0,rm )) , kAkW 1,∞ (B+ (0,rm )) .
As a consequence, similar to (9.30), we obtain that
¯
−(Du, Dv)L2 (Ω) ≤ Cδ kf k2L2 (Ω) + δkξm ∂DΛ m 2
u kL2 (B+ (0,rm )) + (p, divv)L2 (Ω) .
To estimate (p, divv)L2 (Ω) , we note that the change of variable x 7→ θm (x)
implies that
Z h
i 2
−1 i
(p, divv)L2 (Ω) = p θm,` Λ ξm Λ (A`j umj ),α ,α ◦ θm ,i dx
ZΩ
(p ◦ θm )Aki θm,`
i 2
Λ (A`j umj ),α ,α ,k dx

= Λ ξm
B+ (0,rm )
Z
2
Λ (Akj umj ),α ,kα dx,

= (p ◦ θm )Λ ξm
B+ (0,rm )
where we have used that Aki θm,`

i = δ`k together with the Piola identity Aki ,k =
0 (due to the assumption that det(Dθm ) = 1) to conclude the last equality.
Since divu = 0, Akj umj ,k = (divu) ◦ θm = 0 in B+ (0, rm ). Therefore,

Z
(p ◦ θm )Λ ξm ξm ,k Λ (Akj umj ),α ,α dx

(p, divv)L2 (Ω) = 2
B (0,r )
Z + m
(p ◦ θm )Λ ξm ξm ,k Λ (Akj ,α umj ) ,α dx

=2
B+ (0,rm )
Z
(p ◦ θm )Λ (ξm ξm ,k ),α Λ (Akj umj ,α ) dx

+2
B (0,r )
Z + m
(p ◦ θm )Λ ξm ξm ,k ∂α ( Λ , Akj umj ,α ) dx

+2
B (0,r )
Z + m
(p ◦ θm )Λ ξm ξm ,k Akj ,α (Λ umj ),α ) dx

+2
B (0,r )
Z + m
(p ◦ θm )Λ ξm ξm ,k Akj (Λ umj ),αα dx.

+2
B+ (0,rm )
Using (5.46) again to estimate the third term on the right-hand side, by
(9.26) and Young’s inequality we find that
h i
¯
(p, divv)L2 (Ω) ≤ CkpkL2 (Ω) kukH 1 (Ω) + kξm ∂DΛ um
k 2
L (Ω)
¯
≤ Cδ kf k2L2 (Ω) + δkξm ∂DΛ m 2
u kL2 (Ω) ; (9.38)
thus
¯
−(Du, Dv)L2 (Ω) ≤ Cδ kf k2L2 (Ω) + δkξm ∂DΛ m 2
u kL2 (B+ (0,rm )) .
Now we turn our attention to the term which contributes the energy.
First the change of variable x 7→ θm (x) implies that
Z
(Du, Dv)L2 (Ω) = Ark umi ,r Ask v i ,s dx
B+ (0,rm )
Z h i
= Ark umi ,r Ask θm,` i 2
Λ ξm Λ (A`j umj ),α ,α ,s dx.
B+ (0,rm )
Integrating by parts in xα , by the property of convolution we obtain that

Z
2
− (Du, Dv)L2 (Ω) = ξm Λ (Ark Ask θm,`
i
umi ,r ),α Λ (A`j umj ,s ),α dx
B+ (0,rm )
Z h i
2
+ ξm Λ (Ark Ask θm,`
i
umi ,r )Λ (A`j umj ,α ),α +R1 dx,
B+ (0,rm )
where R1 consists of products of first derivatives of um , and satisfies

Z
|R1 |dx ≤ Ckuk2H 1 (Ω) ≤ Ckf k2L2 (Ω) .
B+ (0,rm )
Since Λ (A`j umj ,α ),α = ∂α A`j (Λ umj ),α + ∂α Λ , A`j umj ,α , by (5.46),

Z
2
ξm Λ (Ark Ask θm,`
i
umi ,r )Λ (A`j umj ,α ),α dx
B+ (0,rm )
h i
≤ CkukH 1 (Ω) kukH 1 (Ω) + kξm ∂DΛ ¯ m
u kL2 (B+ (0,rm ))
¯
≤ Cδ kf k2L2 (Ω) + δkξm ∂DΛum 2
kL2 (B+ (0,rm )) .
Moreover,
Λ (Ark Ask θm,`

i
umi ,r ),α Λ (A`j umj ,s ),α
h i
= Ark Ask θm,ì i
(Λ umi ),rα +∂α Λ , Ark Ask θm,`
mi i
u ,r +(Ark Ask θm,` ),α (Λ umi ),r

h i
× A`j (Λ umj ),sα +∂α Λ , A`j umj ,s +A`j ,α (Λ umj ),s ;

thus by (5.46) again,

Z
2
ξm i
Λ (Ark Ask θm,` umi ,r ),α Λ (A`j umj ,s ),α dx
B+ (0,rm )
Z
2 r s
≥ ξm Ak Ak (Λ umj ),rα (Λ umj ),sα dx − Ckuk2H 1 (Ω)
B+ (0,rm )
¯
− CkukH 1 (Ω) kξm ∂DΛum
kL2 (B+ (0,rm ))
¯
≥ (λ − δ)kξm ∂DΛum 2
kL2 (B+ (0,rm )) − Cδ kf k2L2 (Ω) ,
here we use the fact that Ark Ask is positive definite whose smallest eigen-
value is bigger than some positive constant λ, as in the proof of the elliptic
regularity. Therefore, by choosing δ > 0 small enough, we conclude that
¯
kξm ∂DΛ m
u kL2 (B+ (0,rm )) ≤ Ckf kL2 (Ω) ,

where C = C kDζm kL∞ (B+ (0,rm )) , kDθm kL∞ (B+ (0,rm )) , kAkW 1,∞ (B+ (0,rm )) .
The bound on the right-hand side is independent of , so we may pass to
the weak limit (of a subsequence) and obtain that
¯
kξm ∂D(u ◦ θm )kL2 (B+ (0,rm )) ≤ Ckf kL2 (Ω) , (9.39)
Step 3. Boundary estimates for p ◦ θm .

We first assume that Ω = Rn+ to illustrate the idea. Letting ϕ = ψ,α in
(9.27), we find that
(Du, Dψ,α )L2 (Ω) + (p, divψ,α )L2 (Ω) = (f, ψ,α )L2 (Ω) .
Integrating by parts in xα , we find that
(p,α , divψ)L2 (Ω) = −(f, ψ,α )L2 (Ω) − (Du,α , Dψ)L2 (Ω) . (9.40)
Define a bounded functional T : H01 (Ω) → R by
T (ψ) = −(f, ψ,α )L2 (Ω) − (Du,α , Dψ)L2 (Ω) .
By (9.40), we know that T (ψ) = 0 for all ψ ∈ W 1 ; thus the Lagrange multi-
2
Z r ∈ L (Ω)/R such that T (ψ) = (r, divψ)L2 (Ω) .
plier lemma provides a unique
On the other hand, since p,α dx = 0, r = p,α ; thus
Ω
h i
¯
kp,α kL2 (Ω) ≤ CkT kB(H01 (Ω;Rn ),R) ≤ C kf kL2 (Ω) + k∂Duk L (Ω) ≤ Ckf kL2 (Ω) .
2
Now suppose that Ω is a general bounded C 2 -domain. Define pm = p◦θm

−1
and f m = f ◦ θm . Let ϕi = θm,`
i ξm Λ (A`j ψ j ),α ◦ θm be a test function in
(9.27) for some ψ ∈ H01 (Ω; Rn ). Making a change of variable x 7→ θm (x), by
Aki θm,`
i = δ`k and the Piola identity Aki ,k = 0 we find that
− pm , (ξm Λ (Akj ψ j ),α ),k L2 (Ω) = − pm , Aki (θm,ì

ξm Λ (A`j ψ j ),α ),k L2 (B+ (0,rm ))

= Ajr umi ,j , Akr (θm,`

i
ξm Λ (A`j ψ j ),α ),k L2 (B+ (0,rm ))

− f m , θm,`
i
ξm Λ (A`j ψ j ),α L2 (B+ (0,rm )) .

Commuting derivatives and convolution operators about multiplication with

the A matrices, the identity above can be written as
Λ (ξm pm ),α , Akj ψ j ,k L2 (Ω) − Λ (ξm ,k pm ), (Akj ψ j ),α L2 (Ω)

= Λ (Akr Ajr (θm,ì

ξm ),k umi ),j , (A`j ψ j ),α L2 (B+ (0,rm ))

− A`j ,k Λ (Akr Ajr θm,`

i
ξm umi ,j ),α , ψ j L2 (B+ (0,rm ))

(9.41)
− A`j Λ (Akr Ajr θm,ì
ξm umi ,j ),α , ψ j ,k L2 (B+ (0,rm ))

i
ξm f m ), (A`j ψ j ),α L2 (B+ (0,rm )) .

− Λ (θm,`
The identity above motivates us to define the linear functional T given by
T (ϕ) = Λ (ξm ,k pm ), (Akj ψ j ),α L2 (Ω)

+ Λ (Akr Ajr (θm,`

i
ξm ),k umi ),j , (A`j ψ j ),α L2 (B+ (0,rm ))

− A`j ,k Λ (Akr Ajr θm,`

i
ξm umi ,j ),α , ψ j L2 (B+ (0,rm ))

− A`j Λ (Akr Ajr θm,`

i
ξm umi ,j ),α , ψ j ,k L2 (B+ (0,rm ))

i
ξm f m ), (A`j ψ j ),α L2 (B+ (0,rm )) ,

− Λ (θm,`
where ψ on the right-hand side is the same as ϕ ◦ θm . We note that T :

H01 (Ω; Rn ) → R is bounded since
kT kB(H01 (Ω;Rn ),R)

h i
¯ m kL2 (Ω) + kf kL2 (Ω) ≤ Ckf kL2 (Ω)
≤ C kpkL2 (Ω) + kukH 1 (Ω) + kξm ∂Du

for some constant C = C kθm kW 2,∞ (B+ (0,rm )) , kAkW 1,∞ (B+ (0,rm )) . More-
over, T ϕ = 0 for all ϕ ∈ W 1 . Therefore, by the Lagrange multiplier lemma,
there exists a unique r ∈ L2 (Ω)/R such that T (ϕ) = (r, divϕ)L2 (Ω) , and
krkL2 (Ω) ≤ kT kB(H01 (Ω;Rn ),R) ≤ Ckf kL2 (Ω) . (9.42)
On the other hand, by the property of the convolution,

Z Z Z
m −1 m
Λ (ξm p ),α ◦ θm dx = Λ (ξm p ),α dx = (ξm pm ),α dx = 0
Ω B+ (0,rm ) B+ (0,rm )
−1 ∈ L2 (Ω)/R. Since for ϕ = ψ ◦ θ −1 ,

which implies that Λ (ξm pm ),α ◦ θm m
−1
Λ (ξm pm ),α , Aki ψ i ,k = Λ (ξm pm ),α ◦ θm

L2 (Ω)
, divϕ L2 (Ω) ,
the unique existence of r and (9.41) then imply that
−1
r = Λ (ξm pm ),α ◦ θm .
Therefore, (9.42) implies that

by det(Dθm ) = 1 −1
kΛ (ξm pm ),α kL2 (Ω) = kΛ (ξm pm ),α ◦ θm kL2 (Ω) ≤ Ckf kL2 (Ω) .
Since the bound on the right-hand side is independent of , we may pass to

the limit and obtain that
¯ ◦ θm )kL2 (B (0,r )) ≤ Ckf kL2 (Ω) .

kξm ∂(p (9.43)
+ m
Step 4. The estimates for ξm ∂N DΛ (u ◦ θm ) and ξm ∂N (p ◦ θm ).

We first assume that Ω = R3+ to illustrate the idea. Note that (9.37)
implies that
−u1 ,11 −u1 ,22 −u1 ,33 +p,1 = f 1 a.e. in Ω, (9.44a)

−u2 ,11 −u2 ,22 −u2 ,33 +p,2 = f 2 a.e. in Ω, (9.44b)
−u3 ,11 −u3 ,22 −u3 ,33 +p,3 = f 3 a.e. in Ω. (9.44c)
In (9.44), four functions u1 ,33 , u2 ,33 , u3 ,33 and p,3 have unknown regular-
ity; thus an additional equation is required for determining the regularity of
(u,33 , p,3 ). This additional equation comes from the divergence-free condi-
tion: differentiating divu = 0 in x3 , we find that
u1 ,13 +u2 ,23 +u3 ,33 = 0. (9.45)
Equations (9.44) and (9.45) together then form a system of four equations
with four unknowns (or to be more precise, four functions with unknown
regularity), and one can solve for (u,33 , p,3 ) and obtain regularity informa-
tion.
Solving for u1 ,33 and u2 ,33 from (9.44a,b), estimates (9.33), (9.36), (9.39)
and (9.43) imply that u1 ,33 and u2 ,33 both belong to L2 (Ω), and u1 , u2 satisfy
ku1 kH 2 (Ω) + ku2 kH 2 (Ω) ≤ Ckf kL2 (Ω) . (9.46)
Moreover, (9.45) implies that u3 ,33 = −u1 ,13 −u2 ,23 ; thus u3 satisfies
ku3 kH 2 (Ω) ≤ Ckf kL2 (Ω) . (9.47)
Finally, (9.44c) implies that p,3 = f 3 + u3 ,11 +u3 ,22 +u3 ,33 ; thus
h i
kp,3 kL2 (Ω) ≤ C kf kL2 (Ω) + kukH 2 (Ω) ≤ Ckf kL2 (Ω) ,
which, together with (9.36) and (9.43), further implies that
kpkH 1 (Ω) ≤ Ckf kL2 (Ω) .
Now suppose that Ω is a general bounded C 2 -domain. The change of

variable x 7→ θm (x) allows us to rewrite −∆u + Dp = f as
−(Ak` Aj` umi ,j ),k +Aki pm ,k = f i ◦ θm in B+ (0, rm );
thus rearranging terms suggests that
−An` An` umi ,nn+Ani pm ,n = f i ◦ θm −Aαi q,α +Ak` Aj`,k umi ,j + Ak` Aj` umi ,jk .
X
(j,k)6=(n,n)
Let F i be the right-hand side. Then by (9.39) and (9.43),
kξm F kL2 (B+ (0,rm )) ≤ Ckf kL2 (Ω) . (9.48)

As in the proof of the elliptic regularity in Section 5.9, we observe that

An` An` ≥ Ann = eT
n Aen ≥ λ > 0. Therefore, for i = 1, · · · , n,
Ani m Fi
−umi ,nn + p ,n = in B+ (0, rm ). (9.49)
An` An` An` An`
On the other hand, by Aki umi ,k = 0 in B+ (0, rm ),
Akα umα ,k +Aβn umn ,β +Ann umn ,n = 0 in B+ (0, rm );
thus rearranging terms implies that

h i
Ann umn ,n +Anα umα ,n = − Aβα umα ,β +Aγn umn ,γ .
Differentiating the identity above in xn , we find that
Ann umn ,nn +Anα umα ,nn

h i (9.50)
= − Aβα umα ,β +Aγn umn ,γ ,n −Ann,n umn ,n −Anα,n umα ,n .
Let G be the right-hand side of (9.50). Then by (9.39),
kξm GkL2 (B+ (0,rm )) ≤ Ckf kL2 (Ω) . (9.51)
Equations (9.49) and (9.50) form an (n + 1) × (n + 1) system
(An` An` )−1 An1 um1 ,nn (An` An` )−1 F 1

    
−1 0 ··· 0
.. .. .. ..
(An` An` )−1 An2
    
 0 . 0 .  .   . 
.. .. .. ..
    
.. = .
.
 
 . 0 0 .
 .   . 
n n −1 n mn (A` A` )−1 F n
n n
    
 0 ··· 0 −1 (A` A` ) An   u ,nn   
An1 ··· ··· Ann 0 p m ,n G
We note that the determinant of the (n+1)×(n+1) matrix on the left-hand

side is (−1)n ; thus (um1 ,nn , · · · , umn ,nn , pm ,n ) can be uniquely determined.
Moreover, since Ak` ∈ W 1,∞ (B+ (0, rm )), by (9.48) and (9.51) we obtain that
n
X
kξm D2 um kL2 (B+ (0,rm )) + kξm Dpm kL2 (B+ (0,rm )) ≤ Ckf kL2 (Ω) . (9.52)
k=1
Combining (9.33), (9.36) and (9.52) together, we finally conclude that (9.29)
for the case m = 1.
Remark 9.20. As suggested by (9.31), the convolution-by-layers opera-

tor Λ has to act on a divergence-free vector in order to obtain the esti-
mate (9.32), given that p ∈ L2 (Ω). Therefore, in the proof of boundary
2 (Λ umj ),

estimates of Theorem 9.19, instead of using v = Λ ξm α ,α ◦
−1 as a test function (the test function from Section 5.9), we use v i =
θm
i 2 Λ (A` umj ),
−1
θm,` Λ ξm j α ,α ◦ θm as a test function in (9.27), because the
vector w` ≡ A`j umj is divergence-free; that is,
divw = (A`j umj ),` = A`j umj ,` = (divu) ◦ θm = 0.
This is essential for estimating (p, divv)L2 (Ω) .
Now suppose that (9.22b) is replaced by divu = g for some g ∈ H 1 (Ω),

and the boundary condition (9.22c) is replaced by u = h on ∂Ω × (0, T ) for
some h ∈ H 1.5 (∂Ω) satisfying
Z Z
g dx = h · NdS .
Ω ∂Ω
By Corollary 8.5, there exists a function w ∈ H 2 (Ω) satisfying divw = g,

w = h on ∂Ω, and kwkH 2 (Ω) ≤ C kgkH 1 (Ω) +khkH 1.5 (∂Ω) . Letting v = u−w,
we find that v satisfies (9.22) except that the external force is replaced by
f − ∆w, and this process leads to the estimate
h i
kvkH 2 (Ω) ≤ Ckf − ∆wkL2 (Ω) ≤ C kf kL2 (Ω) + kgkH 1 (Ω) + khkH 1.5 (∂Ω) .
We have the following useful generalization:
Theorem 9.21. Suppose that ∂Ω is of class C m+1 for some m ∈ N. Then

for any f ∈ H m−1 (Ω), g ∈ H m (Ω) and h ∈ H m+0.5 (∂Ω) satisfying
Z Z
g dx = h · NdS,
Ω ∂Ω
there exists a unique solution (u, p) ∈ W m+1 × H m (Ω)/R to the steady

state Stokes problem
−∆u + Dp = f in Ω × (0, T ) ,
divu = g in Ω × (0, T ) ,
u=h on ∂Ω × (0, T ) .
Moreover, (u, p) satisfies

h i
kukH m+1 (Ω) + kpkH m (Ω) ≤ C kf kH m−1 (Ω) + kgkH m (Ω) + khkH m+0.5 (∂Ω) .
Corollary 9.22. Let ∂Ω be of class C m for some m ∈ N. Then
H m (Ω; Rn ) ∩ H01 (Ω; Rn ) = W m ⊕ (H m (Ω; Rn ) ∩ G), (9.54)

where G = u ∈ H01 (Ω; Rn ) − ∆u = Dp for some p ∈ L2 (Ω)/R .

Proof. Let E and F be the left and right-hand side space of (9.54), respec-
tively. We prove that F is closed and dense in E to conclude the corollary.
Let un ∈ F and un → u in H m (Ω; Rn ) ∩ H01 (Ω). By the definition of F ,
un = vn + wn with vn ∈ W m and −∆wn = Dqn for some qn ∈ L2 (Ω)/R.
Let fn = −∆un = −∆vn + Dqn . Since fn ∈ H m−2 (Ω), and (vn , qn ) is
the solution to the homogeneous Stokes problem with forcing fn , (vn , qn ) ∈
W m × (H m−1 (Ω)/R) satisfying
kvn kH m (Ω) + kqn kH m−1 (Ω) ≤ Ckfn kH m−2 (Ω) .
Since un → u in H m (Ω; Rn ) ∩ H01 (Ω), fn = −∆un is uniformly bounded

in H m−2 (Ω). Therefore, vn is uniformly bounded in H m (Ω; Rn ) and qn is
uniformly bounded in H m−1 (Ω)/R; thus there exists a subsequence nj of n
such that vnj * v in H m (Ω; Rn ) and qnj * q in H m−1 (Ω)/R. On the other
hand, fn → −∆u in H m−2 (Ω; Rn ), and −∆un = −∆vn + Dqn implies that
−∆u = −∆v + Dq. This proves that F is closed.
To prove the density, let ` ∈ E 0 and `ϕ = 0 for all ϕ ∈ F , and we
show that ` ≡ 0. Since E is dense in H01 (Ω; Rn ), ` has a unique extension
`˜ ∈ H −1 (Ω; Rn ) and `˜ vanishes on W 1 ⊕ G. In particular, `˜ vanishes on W 1 ;
thus Lemma 9.17 implies that
˜ = (p, divϕ)L2 (Ω)

`ϕ ∀ ϕ ∈ H01 (Ω; Rn )
for some p ∈ L2 (Ω)/R. Let z ∈ H01 (Ω; Rn ) be the weak solution of −∆z =
Dp. Then z ∈ G, and by `˜ vanishes on G, we find that for all w ∈ G,
˜ = 0.
hz, −∆wiH01 (Ω) = h−∆z, wiH01 (Ω) = (p, divw)L2 (Ω) = `w
In other words, (divz, q)L2 (Ω) = hz, DqiH01 (Ω) = 0 for all q ∈ L2 (Ω)/R; thus
divz = 0. Therefore, z ∈ W 1 ∩ G = {0} which implies ` ≡ 0.
An orthonormal basis of W 1
Let P denote the Leray projector. The Stokes operator A is defined by

A = −P∆. For f ∈ L2 (Ω; Rn ), we use the notation u = A−1 f to indicate
that u solves
−∆u + Dp = f in Ω,
divu = 0 in Ω,
u=0 on ∂Ω .
By the regularity theory of the Stokes problem, A−1 : W 1 → W 3 is bounded;

thus the Rellich theorem implies that A−1 : W 1 → W 1 is compact. More-
over, suppose that f, g ∈ W 1 , and u = A−1 f , v = A−1 g. Then for some
p ∈ H 2 (Ω),
(A−1 f, g)W 1 = (Du, Dg)L2 (Ω) = (−∆u, g)L2 (Ω)

(9.55)
= (−∆u + Dp, g)L2 (Ω) = (f, g)L2 (Ω) = (f, A−1 g)W 1 .
Therefore, A−1 : W 1 → W 1 is symmetric. By Theorem B.38, there exists

a sequence {ẽj }∞ 1 −1
j=1 ⊆ W , a collection of eigenfunctions of A , such that
{ẽj }∞ 1
j=1 forms an orthonomal basis of W . By the regularity theory for the
Stokes problem, ẽj ∈ W k for all k ∈ N.
Suppose that µj is the eigenvalue associated to ẽj . By (9.55),
µj = µj (ẽj , ẽj )W 1 = (A−1 ẽj , ẽj )W 1 = (ẽj , ẽj )L2 (Ω) > 0;
1/2
thus A−1 is a positive operator on W 1 . Let λj = µ−1
j , and define ej = λj ẽj .
Then {ej }∞ 1 2
j=1 is an orthogonal basis in W which is orthonormal in L (Ω).
Moreover,
Aej = λj ej , λj → ∞ as j → ∞,
or there exists pj ∈ H k (Ω) (for all k) such that
−∆ej + Dpj = λj ej λj → ∞ as j → ∞.
In the following discussion, whenever a basis of W 1 is used, it is this orthog-

onal basis we just construct.
9.4. INCOMPRESSIBLE EULER WITH FREE-BOUNDARY 385
Remark 9.23. Similar to the argument above, A−1 : L2 (Ω; Rn ) → L2 (Ω; Rn )

is compact. Moreover, if f, g ∈ L2 (Ω; Rn ), and u = A−1 f , v = A−1 g, then
for some p, q ∈ H 1 (Ω),
(A−1 f, g)L2 (Ω) = (u, −∆v + Dq)L2 (Ω) = (u, −∆v)L2 (Ω)
= (−∆u, v)L2 (Ω) = (−∆u + Dp, v)L2 (Ω) = (f, A−1 g)L2 (Ω)
which implies that A−1 : L2 (Ω; Rn ) → L2 (Ω; Rn ) is symmetric. Therefore, a

collection of eigenfunctions of A−1 in L2 (Ω; Rn ) constitutes an orthonormal
basis {ēj }∞ 2 n 1
j=1 of L (Ω; R ). However, unlike the case of a basis of W , the
associated eigenvalue of ēj could be zero; that is, there exists ēj 6= 0 such
that A−1 ēj = 0. In this case, there exists pj ∈ H 1 (Ω) such that
Dpj = ēj .
Those eigenfunctions associated with non-zero eigenvalues are also eigen-

functions in W 1 ; thus we establish the Hodge decomposition
L2 (Ω; Rn ) = W 1 ⊕L2 (Ω) Dp p ∈ H 1 (Ω) ,

where the closure of W 1 is taken under the L2 -norm.
9.4 Incompressible Euler with free-boundary

9.4.1 The Eulerian description
For 0 ≤ t ≤ T , the evolution of a three-dimensional incompressible fluid with

a moving free-surface is modeled by the incompressible Euler equations:
ut + u · Du + Dp = 0 in Ω(t) , (9.56a)
divu = 0 in Ω(t) , (9.56b)
p=0 on Γ(t) , (9.56c)
V(Γ(t)) = u · n (9.56d)
u = u0 on Ω(0) , (9.56e)
Ω(0) = Ω . (9.56f)
The open subset Ω(t) ⊂ R3 denotes the changing volume occupied by the
fluid, Γ(t) := ∂Ω(t) denotes the moving free-surface, V(Γ(t)) denotes normal
velocity of Γ(t), and n(t) denotes the exterior unit normal vector to the
free-surface Γ(t). The vector-field u = (u1 , u2 , u3 ) denotes the Eulerian
velocity field, and p denotes the pressure function. We use the notation
D = (∂1 , ∂2 , ∂3 ) to denote the gradient operator. We have normalized the
equations to have all physical constants equal to 1.
This is a free-boundary partial differential equation to determine the ve-
locity and pressure in the fluid, as well as the location and smoothness of the
a priori unknown free-surface. A recent explosion of interest in the analysis
of the free-boundary incompressible Euler equations, particularly in irrota-
tional form, has produced a number of different methodologies for obtaining
a priori estimates, and the accompanying existence theories have mostly re-
lied on the Nash-Moser iteration to deal with derivative loss in linearized
equations when arbitrary domains are considered, or complex analysis tools
for the irrotational problem with infinite depth.
To avoid the use of local coordinate charts necessary for arbitrary ge-
ometries, for simplicity, we will assume that the initial domain Ω at time
t = 0 is given by
Ω = T2 × (0, 1) , (9.57)
where T2 denotes the 2-torus, which can be thought of as the unit square
(0, 1)2 with periodic boundary conditions. This permits the use of one global
Cartesian coordinate system. We only allow the top boundary
Γ = {x3 = 1}
to move, while the bottom boundary is fixed with boundary condition
u3 = 0 on {x3 = 0} × [0, T ] .
Later, we will explain how to treat the case that the initial domain is an
arbitrary bounded, open subset of R3 with H 4 -class boundary.
Einstein’s summation convention
Repeated Latin indices i, j, k,, etc., are summed from 1 to 3; for example,
2
F,ii := i=1,3 ∂x∂i ∂xi .
P
The Lagrangian description
We transform the system (9.56) into Lagrangian variables. We let η(x, t)

denote the “position” of the fluid particle x at time t. Thus,
∂t η = u ◦ η for t > 0 and η(x, 0) = x
where ◦ denotes composition so that [u ◦ η](x, t) := u(η(x, t), t) . We set
v = u ◦ η (Lagrangian velocity),
q = p ◦ η (Lagrangian pressure),
A = [Dη]−1 (inverse of the deformation tensor),
J = det[Dη] (Jacobian determinant of the deformation tensor),
a = J A (cofactor of the deformation tensor).
Since div u = 0, we have that det Dη = 1, and hence the cofactor matrix of
Dη is equal to [Dη]−1 . Using Einstein’s summation convention, and using
∂F
the notation F,k to denote ∂xk , the kth-partial derivative of F for k = 1, 2, 3,
the Lagrangian version of equations (9.56) is given on the fixed reference
domain Ω by
vti + Aki q,k = 0 in Ω × (0, T ] , (9.58a)

divη v = 0 in Ω × (0, T ] , (9.58b)
q=0 on Γ × (0, T ] , (9.58c)
v3 = 0 on {x3 = 0} × (0, T ] , (9.58d)
(η, v) = (e, u0 ) in Ω × {t = 0} , (9.58e)
where e(x) = x denotes the identity map on Ω. Notice that the free suface
Γ(t) is given by
Γ(t) = η(t)(Γ) .
The Lagrangian divergence divη v = Aji v i ,j . Equation (9.58a) can be written

vector form as
vt + AT Dq = 0 in Ω × (0, T ] , (9.58a’)
where AT denotes the transpose of A. Solutions to (9.58) which are suf-

ficiently smooth to ensure that η(t) are diffeomorphisms, give solutions to
(9.56) via the change of variables indicated above.
9.4.2 The Lagrangian vorticity equation
We make use of the permutation symbol


 1, even permutation of {1, 2, 3},
εijk = −1, odd permutation of {1, 2, 3},
0, otherwise ,

and the basic identity regarding the ith component of the curl of a vector
field u:
(curl u)i = εijk uk ,j .
Defining curlη v = curl u ◦ η, the chain rule shows, by taking the curl of the
Euler equations (9.58a), that
(curlη vt )i = εijk asj vtj ,s = 0. (9.59)
9.4.3 Notation
Differentiation and norms
For integers k ≥ 0 and a smooth, open domain Ω of R3 , we define the Sobolev

space H k (Ω) (H k (Ω; R3 )) to be the completion of C ∞ (Ω) (C ∞ (Ω; R3 )) in
the norm
XZ
kuk2k := |Da u|2 dx,
|a|≤k Ω
where Da denotes all partial derivatives of order a. For real numbers s ≥ 0,

the Sobolev spaces H s (Ω) and the norms k · ks are defined by interpolation.
We will write H s (Ω) instead of H s (Ω; R3 ).
We define the horizontal derivative by ∂¯ = (∂1 , ∂2 ), and define the
Sobolev space H k (Γ) to be the completion of C ∞ (Γ) in the norm
XZ
|u|2 :=
k |∂¯a u|2 dS ,
|a|≤k Γ
where ∂¯a denotes all horizontal partial derivatives of order a, and dS =

dx1 dx2 denotes the ‘surface measure.’ For real s ≥ 0, the Hilbert space
H s (Γ) and the boundary norm |·|s is defined by interpolation. The negative-
order Sobolev spaces H −s (Γ) are defined via duality: for real s ≥ 0, H −s (Γ) :=
[H s (Γ)]0 .
The space of divergence-free vectors on Ω
In order to specify our initial velocity field, we introduce the following sub-
space of H s vector-fields on Ω for s ≥ 0:
Definition 9.24 (H s -class divergence free vectors).
s
Hdiv (Ω) = {u ∈ H s (Ω; R3 ) : u3 = 0 on {x3 = 0}, xh 7→ u(xh , x3 ) periodic
div u = 0} .
9.4.4 Properties of the cofactor matrix a, and a polynomial-

type inequality
Differentiating the inverse matrix A
Using that Dη A = Id, we have the following identities
¯ k = −As ∂η
∂A ¯ r , s Ak , (9.60)
i i r
DAki = −Asi Dη r ,s Akr , (9.61)

∂t Aki = −Asi v r ,s Akr . (9.62)
Relating the cofactor matrix and the unit normal n(t)
With N = (0, 0, 1) the outward unit normal to Γ, we have the identity
n(η) = aT N/|aT N | or ni (η) = a3i /|a3· | .
For α, β = 1, 2, we define the components of the induced metric on Γ(t) by
gαβ (x1 , x2 , t) = η,α (x1 , x2 , t) · η,β (x1 , x2 , t),
√ p
and let g denote det(gαβ ). The metric g is a 2 × 2 matrix defined on Γ.
It is an elementary computation to verify that
√
g = |η,1 ×η,2 | = |a3· | on Γ ,
√
and hence that a3i = gni (η).
√
A3i = J −1 gni (η) on Γ . (9.63)
9.4.5 Horizontal convolution-by-layers and commutation es-

timates
Horizontal convolution-by-layers

With xh = (x1 , x2 ), we define ρ ∈ C0∞ (R2 ) by ρ(xh ) = C exp 1
|xh |2 −1
if
|xh | < 1 and ρ(xh ) = 0 if |xh | ≥ 1; we then select the constant C so that
1 xh
R
R2 ρdxh = 1. We define ρκ (xh ) = κ2 ρ( κ ). It follows that for κ > 0,
0 ≤ ρκ ∈ C0∞ (R2 ) with spt(ρκ ) ⊂ B(0, κ). (Here, spt stands for support.)
We define the operation of horizontal convolution-by-layers as follows:
Z
Λκ f (xh , x3 ) = ρκ (xh − yh )f (yh , x3 )dyh for f ∈ L1 (R2 ) .
R2
By standard properties of convolution, there exists a constant C which
is independent of κ, such that for s ≥ 0,
kΛκ F ks ≤ CkF ks ∀ F ∈ H s (Ω) ,
and
|Λκ F |s ≤ C|F |s ∀ F ∈ H s (Γ) .
Furthermore,
¯ κ F k0 ≤ CkF k0
κk∂Λ ∀ F ∈ L2 (Ω) . (9.64)
Commutation estimates
¯ ∈ L2 (Γ), there is a generic

Lemma 9.25. For f ∈ W 1,∞ (Γ) and g, ∂g
constant C independent of κ such that
¯ − f Λκ ∂g|
|Λκ (f ∂g) ¯ 0 ≤ C|f |W 1,∞ (Γ) |g|0 ,
where W 1,∞ (Γ) denotes the Sobolev space of functions u ∈ L∞ (Γ) with weak
¯ ∈ L∞ (Γ).
derivative ∂u
Proof. We have that

1 xh − yh
Z
¯ − f Λκ ∂g](x)
[Λκ (f ∂g) ¯ = ρ( ¯ h )dyh
)[f (yh ) − f (xh )]∂g(y
B(xh ,κ) κ2 κ
1 ¯ x h − yh
Z
= ∂ρ( )[f (yh ) − f (xh )]g(yh )dyh
B(xh ,κ) κ3 κ
| {z }
I1
Z
− ¯ (yh )g(yh )dyh ,
ρκ (xh − yh )∂f
B(xh ,κ)
| {z }
I2
where we have used integration-by-parts in order to obtain the second equal-

ity. From Morrey’s inequality, for all yh ∈ B(xh , κ),
¯ kL∞ (T2 ) ≤ κ|f |W 1,∞ (Γ) ,

|f (xh ) − f (yh )| ≤ κk∂f
1 ¯ xh
so that with K(xh ) = κ2
|∂ρ( κ )|,
|I1 | ≤ |f |W 1,∞ (Γ) K ∗ |g| ,
so by Young’s inequality for convolution integrals,
|I1 |0 ≤ |f |W 1,∞ (Γ) kKkL1 (T2 ) |g|0 .
Similarly,
|I2 |0 ≤ |f |W 1,∞ (Γ) kρκ kL1 (T2 ) |g|0 .
The assertion is proved, given that
1 ¯ yh
Z Z
kKkL1 (T2 ) = |∂ρ( )|dyh = ¯
|∂ρ(z)|dz < ∞.
2
T2 κ κ T2
Lemma 9.26. For κ > 0, there exists C > 0 independent of κ, such that
1
for any g ∈ H 2 (Ω) and f ∈ H 3 (Ω), we have that
Λκ (f g) − f Λκ g 1 ≤ Cκkgk 1 kf k3 + Cκ 12 kgk0 kf k3 .

2 2
Proof. Let ∆ = Λκ (f g) − f Λκ g. Then, we have that

Z
∆(x) = ρκ (xh − yh )[f (yh , x3 ) − f (xh , x3 )] g(yh , x3 ) dyh .
B(xh ,κ)
Using the fact that H 2 (Ω) is embedded in L∞ (Ω), we have that

Z
|∆(x)| ≤ Cκkf k3 ρκ (xh − yh ) |g(yh , x3 )| dyh ,
B(xh ,κ)
showing that
k∆k0 ≤ Cκkf k3 kΛκ |g|k0

≤ Cκkf k3 kgk0 . (9.65)
Now, for j = 1, 2, 3,
∆,j = Λκ (f g,j ) − f Λκ g,j +Λκ (f,j g) − f,j Λκ g.
The difference between the two first terms of the right-hand side of this
identity can be treated in a similar fashion as (9.65), leading us to:
k∆,j k0 ≤ Cκkf k3 kgk1 + kΛκ (f,j g)k0 + kf,j Λκ gk0

≤ Cκkf k3 kgk1 + kf,j gk0 + kf,j kL∞ (Ω) kΛκ gk0
≤ Cκkf k3 kgk1 + 2kf,j kL∞ (Ω) kgk0
≤ Cκkf k3 kgk1 + Ckf k3 kgk0 . (9.66)
Consequently, we obtain by interpolation from (9.65) and (9.66):

1
k∆k 1 ≤ Cκkf k3 kgk 1 + Cκ 2 kf k3 kgk0 .
2 2
Following the estimate (9.66), we can similarly obtain estimates for

k∆,jk k0 and k∆,jkl k0 , j, k, l = 1, 2, 3, and thus estimate the H 2 (Ω)-norm as
well as the H 3 (Ω)-norm of ∆. Interpolation then yields the following:
3
Lemma 9.27. For κ > 0 and s = 2 or 25 , there exists C > 0 independent
of κ, such that for any g ∈ H s (Ω) and f ∈ H 3 (Ω), we have that
Λκ (f g) − f Λκ g ≤ Cκkgks kf k3 + Cκ 12 kgk 1 kf k3 .

s s− 2
9.4.6 An asymptotically consistent κ-approximation of the

Euler equations
The smoothed flow map ηκ
Definition 9.28 (The horizontally-smoothed flow map ηκ ). Suppose the

Lagrangian flow map η is in C([0, T ], H 3 (Ω)). For κ ∈ (0, κ0 ), we set on
[0, T ]
ηκ = Λ2κ η and vκ = ∂t ηκ .
By choosing T > 0 and κ0 sufficiently small, we can ensure that
Aκ (x, t) = [Dηκ (x, t)]−1
is well-defined on [0, T ]. We then define Jκ = det[Dηκ ], and aκ = Jκ Aκ .

Smoothing the initial data
In order to construct solutions to (9.58), we will introduce an approximation

scheme below, and it will be convenient to smooth the initial velocity field.
Let 0 ≤ %κ ∈ C0∞ (R3 ) denote the standard family of mollifiers with
spt(%κ ) ⊂ B(0, κ), and let EΩ denote the Sobolev extension operator map-
ping H s (Ω) to H s (R3 ) for s ≥ 0.
We set w0κ = %κ EΩ (u0 ), and define
uκ0 = w0κ − Drκ ,
where the scalar function rκ is the solution to the elliptic problem
∆rκ = div w0κ in Ω , (9.67a)

xh 7→ rκ (xh , x3 ) periodic x3 ∈ [0, 1] , (9.67b)
rκ = 0 on Γ , (9.67c)
∂rκ
= (w0κ )3 on {x3 = 0} . (9.67d)
∂x3
4 (Ω), it follows that uκ ∈ H s (Ω) for all s ≥ 4, and
Thus, given u0 ∈ Hdiv 0 div
that uκ0 → u0 in H 4 (Ω) as κ → 0.
The approximate κ-problem
For κ ∈ (0, κ0 ), we consider the following sequence of approximate problems:
vt + ATκ Dq = 0 in Ω × (0, Tκ ] , (9.68a)

divηκ v = 0 in Ω × (0, Tκ ] , (9.68b)
q=0 on Γ × (0, Tκ ] , (9.68c)
v3 = 0 on {x3 = 0} × (0, Tκ ] , (9.68d)
(v, η) = (uκ0 , eκ ) on Ω × {t = 0} , (9.68e)
where the solution η = η(κ) depends on the parameter κ, but for nota-
tional simplification, we do not explicitly write this dependence, and where
divηκ v = (Aκ )ji v i ,j , and eκ = Λ2κ e, where e(x) = x denotes the identity map
on Ω. We refer to the approximation (9.68) as the κ-problem. (Note that
our solution is periodic in the x1 and x2 directions.)
9.4.7 Construction of smooth solutions to κ-approximate Eu-

ler equations
Definition 9.29 (The manifold of volume-preserving embeddings). For s ≥

3, we let
s
Dvol = {η ∈ H s (Ω; R3 ) : η({x3 = 0}) ⊂ {x3 = 0}, xh 7→ η(xh , x3 ) periodic
det Dη = 1 , η −1 ∈ H s (η(Ω); Ω) } .
s is a infinite-dimensional C ∞ Hilbert manifold. We let

For s ≥ 3, Dvol
s denote the tangent bundle over D s , and T T D s the second tangent
T Dvol vol vol
s consist of pairs (η, v) ∈ D s ×H s (Ω)◦η,
bundle. Locally, elements of T Dvol vol div
where
s s
Hdiv (Ω) ◦ η = {u ◦ η : u ∈ Hdiv (Ω)} .
Theorem 9.30. Suppose that for s ≥ 4, uκ0 ∈ H s (Ω) with div u0 = 0.

Then for each κ ∈ (0, κ0 ), there exists a unique solution (η(κ), v(κ)) ∈
C ∞ ([0, Tκ ]; T Dvol
s ) to (9.68) with T > 0 depending on kuκ k an on κ > 0.
κ 0 s
Proof. With ηκ and vκ defined in Definition 9.28, we define
u = v ◦ ηκ−1 , uκ = vκ ◦ ηκ−1 , and p = q ◦ ηκ−1 .
It follows from the chain-rule that (9.68) can be written on the time depen-
dent domain ηκ (t, Ω) as
ut + uκ · Du + Dp = 0 in ηκ (t, Ω) , (9.69a)
div u = 0 in ηκ (t, Ω) , (9.69b)
p=0 on ηκ (t, Γ)] , (9.69c)
(u, η) = (uκ0 , e) on Ω × {t = 0} , (9.69d)
Taking the divergence of (9.69a), we see that p(t) satisfies
−∆p = ui ,j ujκ ,i in ηκ (t, Ω) , (9.70a)

p=0 on ηκ (t, Γ) , (9.70b)
∂p
=0 on {x3 = 0} . (9.70c)
∂N
The estimates for p(t) can be easily obtained by transforming (9.5), set on
the smoothed moving domain ηκ (t, Ω), to an elliptic equation on the fixed
domain Ω. It is important to note that this transformation should not
be made with the map ηκ (t), but rather with a family of diffeomorphisms
which inherits the smoothness of ηκ |Γ . To this end, consider the solution to
∆Φ(t) = 0 in Ω with Φ(t) = ηκ (t) on Γ. It follows that
1 1
kΦ(t)ks+1 ≤ C|ηκ |s+1/2 ≤ C(1 + )|η|s−1/2 ≤ C(1 + )kηks , (9.71)
κ κ
where the first inequality is the standard elliptic estimate for the Dirichlet
problem, the second follows from (9.64), and the third from the trace theo-
rem. For κ and Tκ taken sufficiently small kΦ(t)−eks can be made arbitrarily
small on [0, Tκ ], from which it follows that each such Φ(t) : Ω → ηκ (t, Ω) is
a diffeomorphism.
Next, define the matrix B = [DΦ]−1 and the pressure function Q = p◦Φ.
Using the chain-rule, (9.5) is transformed to
−Bij [Bik Q,k ],j = ui ,j ujκ ,i ◦ Φ

in Ω , (9.72a)
Q=0 on Γ , (9.72b)
Q,k Bik Bi3 = 0 on {x3 } = 0 , (9.72c)
for 0 ≤ t ≤ Tκ . Elliptic estimates (in conjunction with the Sobolev embed-

ding theorem) show that
kQ(t)ks+1 ≤ CP (kΦ(t)ks+1 )kui ,j ujκ ,i ◦Φks−1
where P is a polynomial function of its argument; hence, together with

(9.71), we have the estimate
kQ(t)ks+1 ≤ Cκ P (kη(t)ks ) · kvk2s ,
where the constant Cκ depends on κ > 0 and, in fact, blows-up as κ → 0.

Since p = Q ◦ Φ−1 , we see that
kDp(t) ◦ η(t)ks ≤ Cκ P (kη(t)ks , kvks ) . (9.73)
s by
We define the function F on T Dvol
F(η, v) = −Dp ◦ ηκ ,
and write (9.68a) as the coupled system of first-order ordinary differential

s :
equations written on T Dvol
∂t (η, v) = (v, F) ,
(η, v)|t=0 = (e, u0 ) .
s → T T D s continuously;
According to the estimate (9.73), (v, F) : T Dvol vol
moreover, we proved in establishing the inequality (9.15) that the map η 7→
s to T T D s so that
Dp ◦ η is Lipschiptz continuous from T Dvol vol
s s
(v, F) : T Dvol → T T Dvol is a Lipschitz map.
In fact, this map is C ∞ . Thus, the fundamental theorem of ordinary differ-

ential equations (or Picard iteration) shows that there exists a time Tκ > 0
depending on the initial data (and, of course, κ > 0) such that
(η, v) = (η(κ), v(κ)) ∈ C ∞ ([0, Tκ ]; T Dvol

s
)
is a unique solution of (9.68). Since, by definition Q = 0 on Γ, and
q = Q ◦ Φ−1 ◦ ηκ , (9.74)
it follows that q = 0 on Γ as well.
9.4.8 Asymptotic estimates which are independent of the

smooth parameter κ
According to Theorem 9.30, we have unique solutions to (9.68),
(η(κ), v(κ)) ∈ C ∞ ([0, Tκ ]; T Dvol

s
),
and q(κ) given by (9.74). We will take s ≥ 6, and for notational convenience
we will denote (η(κ), v(κ)) by (η̃, ṽ), and write Ã for [Dη̃]−1 . We use the
notation η̃κ to denote Λ2κ η̃ and set Ãκ = [Dη̃κ ]−1 .
A continuous-in-time energy function appropriate for the asymp-

totic process κ → 0.
Definition 9.31. We set on [0, Tκ ]

√
Eκ (t) = 1 + kΛκ η̃(t)k24.5 + kṽ(t)k24 + kκṽ(t)k24.5 + kṽt (t)k23.5 + k κṽt (t)k24 .
(9.75)
The function Eκ (t) is the higher-order energy function which we will

prove remains bounded on a time-interval which is independent of κ. Given
6 , the E (t) is continuous on [0, T ].
(η̃, ṽ) ∈ T Dvol κ κ
Definition 9.32. We set E(t) = Eκ=0 (t) .
Definition 9.33. We set the constant M0 to be a polynomial function of

E(0) so that
M0 = P (E(0), k curl uκ0 k23.5 ) . (9.76)
Statement of the main result

4 , we obtain the initial pressure func-
Given an initial velocity field u0 ∈ Hdiv
tion p0 as the solution to the elliptic equation
−∆p0 = ui0 ,j uj0 ,i in Ω ,

p0 = 0 on Γ ,
p 0 ,3 = 0 on {x3 = 0} .
4 (Ω) with curl u ∈

Theorem 9.34 (Main Result). Given initial data u0 ∈ Hdiv 0
H 3.5 (Ω) such that
∂p0
− (x) ≥ λ > 0 for x ∈ Γ,
∂x3
there exists a solution to (9.58) verifying
sup E(t) ≤ P (E(0)) .

t∈[0,T ]
5 (Ω) and curl u ∈ H 4.5 (Ω), then the solution is unique.

If u0 ∈ Hdiv 0
Remark 9.35. The same theorem and proof hold in the case that Ω ⊂ R2 .
Below, we will treat the case of a general initial domain Ω.
Remark 9.36. The regularity for the existence theory is not optimal. In
fact, for our domain Ω, all that is necessary to establish existence and
3 (Ω);
uniqueness of solutions to (9.58) is an initial velocity field u0 ∈ Hdiv
nevertheless, the assumptions of Theorem 9.34 allow for the most transpar-
ent proof.
Conventions about constants

6 ,
As noted above, Theorem 9.30 provides us with solutions (η̃, ṽ) ∈ T Dvol
and hence supt∈[0,Tκ ] Eκ (t) is continuous.
We take Tκ > 0 sufficiently small so that, using the fundamental theorem
of calculus, for constants c1 , c2 and t ∈ [0, Tκ ],
−q̃,3 (t) ≥ λ/2 ,

c1 det g̃κ (0) ≤ det g̃κ (t) ≤ c2 det g̃κ (0) on Γ ,
c1 det J˜κ (0) ≤ det J˜κ (t) ≤ c2 det J˜κ (0) in Ω ,
kΛ̃κ η(t)k4 ≤ |Λκ e|4 + 1 , kq̃(t)k4 ≤ kq̃(0)k4 + 1 ,
kṽ(t)k3.5 ≤ ku0 k3.5 + 1 , kṽt (t)k3 ≤ kṽt (0)k3 + 1 .
The right-hand sides appearing in the last three inequalities shall be denoted
by a generic constant C in the estimates that we will perform.
Curl and divergence estimates for η̃, ṽ, and ṽt
Proposition 9.37. For all t ∈ (0, T ), with T ≤ Tκ ,

√
kcurl Λκ η̃(t)k23.5 + kcurl ṽ(t)k23 + kκcurl ṽ(t)k23.5 + k κcurl ṽt (t)k22.5
≤ M0 + C T P ( sup Eκ (t)) . (9.77)
t∈[0,T ]
Proof. By taking the curl of (9.68a), we have that
curlηκ ṽt = 0 .
From (9.59), it follows that (curlη̃κ ṽ)kt = B(Ãκ , Dṽ), where
s
B(Ãκ , Dṽ) = εkji (Ãκ )t j ṽ i ,s = εkij ṽ i ,s (Ãκ )sp ṽκp ,l (Ãκ )lj ;
hence,
Z t
curlη̃κ ṽ(t) = curl uκ0 + B(Ãκ (t0 ), Dṽ(t0 ))dt0 . (9.78)
0
Step 1. Estimate for curl Λκ η̃. Computing the gradient of (9.78) yields
Z t
curlη̃κ Dṽ(t) = D curl uκ0 − ε·ji D(Ãκ )sj ṽ i ,s + DB(Ãκ (t0 ), Dṽ(t0 ))dt0 .
0
Applying the fundamental theorem of calculus once again, shows that

Z t
s
curlη̃κ Dη̃(t) = tD curl uκ0 + ε·ji [(Ãκ )t j Dη̃ i ,s −D(Ãκ )sj ṽ i ,s ]dt0
0
Z tZ t0
+ DB(Ãκ (t00 ), Dṽ(t00 ))dt00 dt0 ,
0 0
and finally that
Z t
s
D curl η̃(t) = tD curl uκ0 − ε·ji (Ãκ )t j (t0 )dt0 Dη̃ i ,s (9.79)
0
Z t Z tZ t0
s
+ ε·ji [(Ãκ )t j Dη̃ i ,s −D(Ãκ )sj ṽ i ,s ]dt0 + DB(Ãκ (t00 ), Dṽ(t00 ))dt00 dt0 .
0 0 0
Using the fact that ∂t (Ãκ )sj = −(Ãκ )sl ṽκl ,p (Ãκ )pj and D(Ãκ )sj = −(Ãκ )sl Dη̃κl ,p (Ãκ )pj ,
we see that
DB(Ã, Dṽ) = −εkji [Dṽ i ,s (Ãκ )sl ṽκl ,p (Ãκ )pj + ṽ i ,s (Ãκ )sl Dṽκl ,p (Ãκ )pj
+ ṽ i ,s ṽκl ,p D((Ãκ )sl (Ãκ )pj )] .
The precise structure of the right-hand side is not very important; rather,
the derivative count is the focus, and as such we write
DB(Ã, D̃v) ∼ D2 ṽ Dṽκ Ãκ Ãκ + D2 ṽκ Dṽ Ãκ Ãκ + D2 η̃κ Dṽ Dṽκ Ãκ Ãκ .
Integrating by parts in time in the last term of the right-hand side of

(9.79), we see that
Z tZ t0 Z tZ t0 h i
DB(Ãκ , Dṽ) dt00 dt0 ∼ − D2 η̃ (Dṽκ Ãκ Ãκ )t + D2 η̃κ (Dṽ Ãκ Ãκ )t dt00 dt0
0 0 0 0
Z tZ t0
+ D2 η̃κ Dṽ Dṽκ Ãκ Ãκ dt00 dt0
0 0
Z th i
+ D2 η̃ Dṽκ Ãκ Ãκ + D2 η̃κ Dṽ Ãκ Ãκ dt0 .
0
Thus, we can write
Z t Z t
0
D curl η̃(t) ∼ tD curl uκ0 + D η̃2
Dṽκ Ãκ Ãκ dt + D2 η̃Dṽκ Ãκ Ãκ dt0
0 0
| {z } | {z }
I1 I2
Z t Z tZ t0
+ D2 η̃κ Dṽ Ãκ Ãκ dt0 + D2 η̃κ Dṽ Dṽκ Ãκ Ãκ dt00 dt0
|0 {z } |0 0
{z }
I3 I4
Z tZ t0 Z tZ t0
+ D2 η̃κ (Dṽ Ãκ Ãκ )t dt00 dt0 + D2 η̃ (Dṽκ Ãκ Ãκ )t dt00 dt0 .
|0 0
{z } |0 0
{z }
I5 I6
Our goal is to estimate kD curl Λκ ηk22.5 , which in turn requires us to

estimate kΛκ Ii k22.5 for i = 1, ..., 6. We begin with i = 1:
Z t
2 2
kΛκ I1 k2.5 ≤ kD Λκ η̃ Dṽκ Ãκ Ãκ dt0 k22.5
0
Z t Z t
+ kΛκ D2 η̃ Dṽκ Ãκ Ãκ dt0 − Dṽκ Ãκ Ãκ dt0 Λκ D2 η̃k22.5
0 0
It is easy to see that

Z t
2
kD Λκ η̃ Dṽκ Ãκ Ãκ dt0 k22.5 ≤ CT P ( sup Eκ (t)) ,
0 t∈[0,T ]
and by Lemma 9.27

Z t Z t
kΛκ D2 η̃ Dṽκ Ãκ Ãκ dt0 − Dṽκ Ãκ Ãκ dt0 Λκ D2 η̃k22.5
0 0
Z t Z t
0 2
2 2 2
≤ Cκ kD η̃k2.5 k 2 2
Dṽκ Ãκ Ãκ dt k3 + CκkD η̃k2 k Dṽκ Ãκ Ãκ dt0 k23
0 0
≤ CT P ( sup Eκ (t))
t∈[0,T ]
The same type of commutation estimate shows that
kΛκ I2 k22.5 ≤ CT P ( sup Eκ (t)) .

t∈[0,T ]
For i = 3, 4, 5, we use that kΛκ Ii k2.5 ≤ kIi k2.5 and as H 2.5 (Ω) is a multi-
plicative algebra, we see that for i = 3, 4, 5,
kΛκ Ii k22.5 ≤ CT P ( sup Eκ (t)) .

t∈[0,T ]
Finally, we consider the case that i = 6:

Z tZ t0
kΛκ I6 k22.5 ≤ k D2 Λκ η̃ (Dṽκ Ãκ Ãκ )t dt00 dt0 k22.5
0 0
Z tZ t0 h i
+k Λκ D2 η̃ (Dṽκ Ãκ Ãκ )t − Dṽκ Ãκ Ãκ D2 Λκ η̃ dt00 dt0 k22.5
0 0 t
≤ CT P ( sup Eκ (t)) ,
t∈[0,T ]
where we have used Lemma 9.27 for the last inequality together with the
1
fact that kκ 2 ṽt k24 is contained in the energy function Eκ (t). Therefore, we
have proven that
kD curl Λκ ηk22.5 ≤ M0 + CT P ( sup Eκ (t)) ,

t∈[0,T ]
and hence with curlη̃κ ṽt = 0, that
k curl Λκ ηk23.5 ≤ M0 + CT P ( sup Eκ (t)) .

t∈[0,T ]
Step 2. Estimate for curl ṽ. From (9.78),

Z t Z t
curl v(t) = curl uκ0 + B(Ãκ (t0 ), Dṽ(t0 ))dt0 − ε·jk ṽ k ,r ∂˜t (Aκ )rj dt0 .
0 0
As H 3 (Ω) is a multiplicative algebra, it follows that on [0, T ],
k curl ṽ(t)k23 ≤ M0 + CT P ( sup Eκ (t)) .

t∈[0,T ]
Similarly,
kκ curl ṽ(t)k23.5 ≤ M0 + CT P ( sup Eκ (t)) .
t∈[0,T ]
Step 3. Estimate for curl ṽt . From (9.78),

Z t
curl ṽt (t) = εikj ∂t (Ãκ )rj (t0 )dt0 ṽti ,r ,
0
from which it follows that on [0, T ],

√
k κ curl ṽt (t)k23 ≤ CT P ( sup Eκ (t)) .
t∈[0,T ]
Proposition 9.38. For all t ∈ (0, T ), with T ≤ Tκ ,

√
kdiv Λκ η̃(t)k23.5 + kdiv ṽ(t)k23 + kκdiv ṽ(t)k23.5 + k κdiv ṽt (t)k23
≤ M0 + C T P ( sup Eκ (t)) . (9.80)
t∈[0,T ]
Proof. Since (Ãκ )ji ṽ i ,j = 0, we see that
(Ãκ )ji Dṽ i ,j = −D(Ãκ )ji ṽ i ,j . (9.81)
Step 1. Estimate for div Λκ η̃. It follows that
[(Ãκ )ji Dη̃ i ,j ]t = ∂t (Ãκ )ji Dη̃ i ,j −D(Ãκ )ji ṽ i ,j
so that
Z t
[(Ãκ )ji Dη̃ i ,j ](t) = ∂t (Ãκ )ji Dη̃ i ,j −D(Ãκ )ji ṽ i ,j dt0 ,
0
and hence
Z t Z t Z t
D div η̃(t) = ∂t (Ãκ )ji Dη̃ i ,j dt0 − D(Ãκ )ji ṽ i ,j dt0 − ∂t (Ãκ )ji dt0 Dη̃ i ,j .
|0 {z } |0 {z } |0 {z }
I1 I2 I3
Thus,
3
X
kD div Λκ η̃(t)k22.5 ≤ kΛκ Ii (t)k22.5 .
i=1
Using Lemma 9.27 in the same fashion as was used for Propostion 9.37, we
see that
3
X
kΛκ Ii (t)k22.5 ≤ CT P ( sup Eκ (t)) ,
i=1 t∈[0,T ]
k div Λκ η̃(t)k23.5 ≤ CT P ( sup Eκ (t)) .

t∈[0,T ]
Step 2. Estimate for div ṽ. From (Ãκ )ji ṽ i ,j = 0, we see that
Z t
div ṽ(t) = − ∂t (Ãκ )ji dt0 ṽ i ,j . (9.82)
0
Hence, it is clear that
k div ṽ(t)k23 + kκ div ṽ(t)k23.5 ≤ CT P ( sup Eκ (t)) .

t∈[0,T ]
Step 3. Estimate for div ṽt . From (9.82),

Z t
div ṽt (t) = −∂t (Ãκ )ji ṽ i ,j − ∂t (Ãκ )ji dt0 ṽti ,j .
0
It follows that
√
k div ṽt (t)k22.5 + k κ div ṽt (t)k23 ≤ CT P ( sup Eκ (t)) .
t∈[0,T ]
Pressure estimates
Letting (Ãκ )ji ∂x∂ j act on (9.68a), for t ∈ [0, Tκ ], the pressure function q(x, t)
satisfies the elliptic equation
h i
−(Ãκ )ji (Ãκ )ki q̃,k ,j = ṽ i ,j (Ãκ )jr ṽκr ,s (Ãκ )si in Ω ,
q̃ = 0 on Γ ,
q̃,k (Ãκ )ki (Ãκ )3i = ṽ · ∂t (Ãκ )3· = 0 on {x3 = 0} .
Using our conventions of Section 9.4.8 concerning the generic constant

C, we have the standard elliptic estimate on [0, Tκ ]
kq̃(t)k4.5 ≤ CkΛκ η̃(t)k4.5 . (9.83)
Similarly by time-differentiating the above elliptic equation for q, elliptic

estimates show that
kq̃t (t)k24 ≤ C kΛκ η̃(t)k24.5 + kṽ(t)k24 .

Technical lemma
Our energy estimates require the use of the following

1 1
Lemma 9.39. Let H 2 (Ω)0 denote the dual space of H 2 (Ω). There exists a
positive constant C such that
1
¯ k
k∂F 1 ≤ C kF k 1 ∀F ∈ H 2 (Ω) .
H 2 (Ω)0 2
Proof. Integrating by parts with respect to the horizontal derivative yields

for all G ∈ H 1 (Ω),
Z Z
¯ G dx = −
∂F ¯ dx ≤ CkF k0 kGk1 ,
F ∂G
Ω Ω
which shows that there exists C > 0 such that
¯ kH 1 (Ω)0 ≤ CkF k0 .
∀F ∈ L2 (Ω), k∂F (9.84)
Interpolating with the obvious inequality
¯ kL2 (Ω) ≤ CkF k1

∀F ∈ H 1 (Ω), k∂F
proves the lemma.
Energy estimates
In this section, we take T ∈ (0, Tκ ).
Proposition 9.40. For t ∈ [0, Tκ ],
|Λκ η̃ 3 (t)|24 + |ṽ 3 (t)|23.5 ≤ M0 + C T P ( sup Eκ (t)) . (9.85)

t∈[0,T ]
Proof. Taking the L2 (Ω) inner-produce ∂¯4 of (9.68a) with ∂¯4 ṽ i yields
1 d ¯4
Z Z
0= k∂ ṽ(t)k0 + ∂ (Ãκ )i q̃,k ∂ ṽ dx + (Ãκ )ki ∂¯4 q̃,k ∂¯4 ṽ i dx +R ,
2 ¯4 k ¯4 i
2 dt } |Ω
| {z {z } |Ω {z }
I1 I2 I3
(9.86)
where R denotes integrals consisting of lower-order terms which can easily
be shown, via the Cauchy-Schwarz inequality, to satisfy
Z T
|R(t)|dt ≤ M0 + C T P ( sup E(t)) .
0 t∈[0,T ]
Using the identity (9.60), we see that

Z
I2 = − (Ãκ )kr ∂¯4 η̃κr ,s (Ãκ )si q̃,k ∂¯4 ṽ i dx + R
ZΩ Z
k ¯4 r
= − (Ãκ )r ∂ ηκ (Ãκ )i q̃,k ∂ ṽ dxh + (Ãκ )kr ∂¯4 ηκr (Ãκ )si q,k ∂¯4 ṽ i ,s dx +R .
3 ¯4 i
|Γ {z } |Ω {z }
I2a I2b
Since q̃ = 0 on Γ, so that q̃,k = q̃,3 , we see that

Z
−I2a = (−q̃,3 )∂¯4 η̃κr (Ãκ )3r ∂¯4 ṽ i (Ãκ )3i dxh ,
Γ
Recalling that η̃κ = Λκ Λκ η̃,

Z
−I2a = (−q̃,3 )∂¯4 Λκ η̃ r (Ãκ )3r ∂¯4 Λκ ṽ i (Ãκ )3i dxh
|Γ {z }
K1
Z h i
+ ∂¯4 Λκ η̃ r Λκ (−q̃,3 ) (Ãκ )3i (Ãκ )3r ∂¯4 ṽ i − (−q̃,3 ) (Ãκ )3i (Ãκ )3r Λκ ∂¯4 ṽ i dxh ,
|Γ {z }
K2
According to Lemma 9.25,

K2 (t) ≤ |∂¯4 Λκ η̃ r |0 Λκ (−q̃,3 ) (Ãκ )3i (Ãκ )3r ∂¯4 ṽ i − (−q̃,3 ) (Ãκ )3i (Ãκ )3r Λκ ∂¯4 ṽ i

0
≤ C|∂¯4 Λκ η̃ r |0 |q̃,3 (Ãκ )3i (Ãκ )3r |W 1,∞ (Γ) |∂¯3 ṽ i |0 .
By the Sobolev embedding theorem,
|q̃,3 (Ãκ )3i (Ãκ )3r |W 1,∞ (Γ) ≤ C|q̃,3 |1.5 |(Ãκ )3i |1.5 |(Ãκ )3r |1.5 ,
so that Z T
K2 (t)dt ≤ C T P ( sup Eκ (t)) .
0 t∈[0,T ]
To study the integral K1 , we define ñκ to be the outward unit normal to

the moving surface η̃κ (t, Γ), given by ñκ = (ãκ )3i /|(ãκ )3· |. From (9.63),
(Ãκ )3· = J˜κ−1

p
g̃κ nκ on Γ ,
where
p
g̃κ = |η̃κ ,1 ×η̃κ ,2 | = |(ãκ )3· | on Γ .
It follows that
Z
K1 (t) = (−q̃,3 )∂¯4 Λκ η̃ · ñκ ∂¯4 Λκ ṽ · ñκ | det g̃κ |J˜κ−2 dxh
Γ
1d
Z
= (−q̃,3 )|∂¯4 Λκ η̃ · ñκ |2 | det g̃κ |J˜κ−2 dxh
2 dt Γ
| {z }
K1a
1 ¯4
Z
− ∂ Λκ η̃ i ∂¯4 Λκ η̃ j ∂t [(ñκ )i (ñκ )j | det g̃κ | J˜κ−2 ]dxh .
Γ 2
| {z }
K1b
By the assumption of Section 9.4.8,

˜−2
∂t [(ñκ )i (ñκ )j | det g̃κ | Jκ ]

≤C,
L∞ (Γ)

Z T
K1b (t)dt ≤ C T P ( sup Eκ (t)) .
0 t∈[0,T ]
Using our assumed bounds for −q̃,3 (t), det g̃κ (t), J˜κ on [0, Tκ ], we see that
Z T
¯4 2
c̄|∂ Λκ η̃(t) · ñκ (t)|0 ≤ K1 (t)dt + M0 + C T P ( sup Eκ (t)) ,
0 t∈[0,T ]
for a constant c̄ which depends on λ, g̃κ (0), J˜κ (0).

Rt
Notice that N = (0, 0, 1) = ñκ (t)− 0 ∂t ñκ (t0 )dt0 , and by our assumptions
in Section 9.4.8, |∂t ñκ (t)|L∞ (Γ) ≤ C, so that with η̃ 3 = η̃ · N ,
Z T
3
c̄|Λκ η̃ (t)|24 ≤ K1 (t)dt + M0 + C T P ( sup Eκ (t)) ,
0 t∈[0,T ]
and hence
Z T
3
c̄|Λκ η̃ (t)|24 ≤− I2a (t)dt + M0 + C T P ( sup Eκ (t)) ,
0 t∈[0,T ]
RT RT
It remains to show that the integrals 0 I2b (t)dt and 0 I3 (t)dt are both
bounded by C T P (supt∈[0,T ] Eκ (t)). Using (9.68b),
Z
I2b (t) = − (Ãκ )kr ∂¯4 η̃κr q̃,k ṽ i ,s ∂¯4 (Ãκ )si dx + R
Ω
¯4
≤ Ck∂ η̃κ (t)k 1 k∂¯4 Ãκ (t)k 1 +R
2 H 2 (Ω)0
≤ Ck∂¯4 η̃κ (t)k 1 k∂¯3 Ãκ (t)k 1 +R

2 H 2 (Ω)
≤ C sup E(t) + R ,
t∈[0,T ]
where we have used Lemma 9.39 for the second inequality.

Finally,
Z Z
I3 (t) = − ∂¯4 q̃ ∂¯4 ṽ i ,k (Ãκ )ki dx = ∂¯4 q̃ ṽ i ,k ∂¯4 (Ãκ )ki dx + R
Ω Ω
¯3
≤ Ck∂ q̃(t)k 1 k∂¯4 Ãκ (t)k 1 +R
2 H 2 (Ω)0
≤ C sup E(t) + R ,
t∈[0,T ]
where we have used the pressure estimate (9.83) and Lemma 9.39 for the
last inequality.
Summing the estimates for I1 , I2 , I3 and integrating (9.86) from 0 to
T , we obtain the inequality,
|η̃ 3 |24 + k∂¯4 ṽ(t)k24 ≤ M0 + C T P ( sup Eκ (t)) .

sup
t∈[0,T ] t∈[0,T ]
According to Proposition 9.38,
sup k div ṽ(t)k23 ≤ M0 + C T P ( sup Eκ (t)) ,

t∈[0,T ] t∈[0,T ]
sup k∂¯4 div ṽ(t)k2H 1 (Ω)0 ≤ M0 + C T P ( sup Eκ (t)) .

t∈[0,T ] t∈[0,T ]
Hence, the normal trace estimate (9.1) shows that
|η̃ 3 (t)|24 + |ṽ 3 (t)|23.5 ≤ M0 + C T P ( sup Eκ (t)) .

sup (9.87)
t∈[0,T ] t∈[0,T ]

Combining Proposition 9.40 with the curl estimates in Proposition 9.37

and the divergence estimates in Proposition 9.38 for η(t) and v(t) and using
(??) provides the following
kΛκ η̃(t)k24.5 + kṽ(t)k24 ≤ M0 + C T P ( sup Eκ (t)) . (9.88)

t∈[0,T ]
Since ṽt = −ÃTκ Dq̃, we also obtain
kṽt (t)k23.5 ≤ M0 + C T P ( sup Eκ (t)) . (9.89)

t∈[0,T ]
Proposition 9.43. For t ∈ [0, Tκ ] and α = 1, 2,

√
| κṽtα (t)|23.5 + |κṽ α (t)|24 ≤ M0 + C T P ( sup Eκ (t)) . (9.90)
t∈[0,T ]
Proof. Multiplying (9.68a) by η̃κj , we find the identity
ṽt · η̃κ ,k = −q̃,k for k = 1, 2, 3 .
It follows that
¯ κ = −∂¯q̃ = 0 on Γ ,
ṽt · ∂η
so that the tangential component (with respect to the moving boundary

η̃κ (t, Γ)) of ṽt vanishes on Γ. Hence,
∂¯3 ṽt · ∂¯η̃κ = −ṽt · ∂¯4 η̃κ − 3∂¯2 ṽt · ∂¯2 η̃κ − 3∂ṽ
¯ t · ∂¯3 η̃κ .
Thus,
|∂¯3 ṽt · ∂¯η̃κ |0 ≤ C|ṽt |2.5 |η̃κ |4 ≤ C|η̃κ |4 ,
so that using the fundamental theorem of calculus, together with the fact
that ηκ ,α (0) is proportional to Tα for α = 1, 2, we find that
|ṽtα |3 ≤ C|η̃κ |4 + C T |ṽt |3 , (9.91)
Similarly,
|∂¯4 ṽt · ∂¯η̃κ |0 ≤ C|ṽt |3 |η̃κ |5 ≤ C|η̃κ |5 ,
|ṽtα |4 ≤ C|η̃|5 + C T |ṽt |4 . (9.92)
Interpolation between (9.91) and (9.92) yields
|ṽtα |3.5 ≤ C|η̃κ |4.5 + C T |ṽt |3.5 .
Interpolation between the inequalities |η̃κ |4 ≤ C|Λκ η̃|4 and |η̃κ |5 ≤ Cκ |Λκ η̃|4
√
shows that κ|η̃κ |4.5 ≤ C|Λκ η̃|4 . It thus follows from (9.88) and (9.89) that
√
κ|ṽtα (t)|3.5 ≤ M0 + C T P ( sup Eκ (t)) .
t∈[0,T ]
A similar argument shows that
κ|ṽ α (t)|4 ≤ M0 + C T P ( sup Eκ (t)) .

t∈[0,T ]
Combining the estimate (9.90) together with the curl estimates in Propo-
√
sition 9.37 and divergence estimates in Proposition 9.38 for κṽ(t) and κṽt (t)
proves the following
Proposition 9.44. For t ∈ [0, Tκ ] and α = 1, 2,

√
k κṽt (t)k24 + kκṽ(t)k24.5 ≤ M0 + C T P ( sup Eκ (t)) . (9.93)
t∈[0,T ]
9.4.9 Proof of the Main Theorem

9.4.10 Time of existence and bounds independent of κ and
existence of solutions to (9.58)
Summing the inequalities provided by the above Propositions, we find that
sup Eκ (t) ≤ M0 + C T P ( sup Eκ (t)) .

t∈[0,T ] t∈[0,T ]
A standard continuation argument then provides us with a time of ex-

istence T1 independent of κ and an estimate on (0, T1 ) independent of κ of
the type:
sup Eκ (t) ≤ 2M0 . (9.94)
t∈[0,T1 ]
In particular, our sequence of solutions (η̃, ṽ) satisfy the κ-independent

bound (9.94) on the κ-independent time-interval (0, T1 ).
9.4.11 The limit as κ → 0
By the κ-independent estimate (9.94), there exists a subsequence of {ṽt , Ãκ , Dq̃}
which converges uniformly to (vt , A, Dq) where A = [Dη]−1 , and η =
Rt
e + 0 vdt0 . Standard arguments show that (η, v) solve (9.58), and that
sup E(t) ≤ P (E(0)) .

t∈[0,T1 ]
Uniqueness
Suppose that on [0, T1 ], (η 1 , v 1 , q 1 ) and (η 2 , v 2 , q 2 ) are both solutions of

5 (Ω) and curl u ∈ H 4.5 (Ω), and with both
(9.58) with initial data u0 ∈ Hdiv 0
q1 and q2 satisfying the stability condition −∂qi /∂N > 0 on Γ for i = 1, 2.

Setting
Eη (t) = 1 + kη(t)k25.5 + kv(t)k25 ,
by the method of the previous section (with κ = 0), we infer that both Eη1 (t)
and Eη2 (t) are bounded by a constant M0 depending on the data u0 and Γ
on a time interval 0 ≤ t ≤ T1 for T1 small enough.
Let
w := v 1 − v 2 , r := q 1 − q 2 , and ξ := η 1 − η 2 .
Then (ξ, w, r) satisfies

Z t
ξ= w in Ω × (0, T ] , (9.95a)
0
∂t wi + (a1 )ki r,k = (a − a1 )ki q 2 ,k
2
in Ω × (0, T ] , (9.95b)
i
(a1 )ji wi ,j = (a2 − a1 )ji v 2 ,j in Ω × (0, T ] , (9.95c)
r=0 on Γ × (0, T ] , (9.95d)
(ξ, w) = (0, 0) on Ω × {t = 0} . (9.95e)
We set
E(t) = 1 + kξ(t)k24.5 + kw(t)k24 .
We will show that E(t) = 0, which shows that w = 0. We follow the

identical analysis as in the previous section, and estimate the new error
terms, arising from the difference of two solutions, using the additional space
regularity in our assumptions. With (9.95e), we find that supt∈[0,T ] E(t) ≤
C T P (supt∈[0,T ] E(t)).
Optimal regularity for initial data
We smoothed our initial data u0 in order to construct solutions to the κ-

approximation (9.68). Having obtained solutions which depend only on
E(0), a standard density argument shows that the initial data needs only
to satisfy E(0) < ∞. All assumptions on from Section 9.4.8 can now be
verified by the fundamental theorem of calculus, and taking T1 even small
if necessary.
Appendix A
A Brief Review on Analysis
A.1 Metric spaces
Let X be a set, and d : X × X → R be a real-valued function. (X, d) is

called a metric space if d satisfies
(1) d(x, y) ≥ 0 . d(x, y) = 0 if and only if x = y .
(2) d(x, y) = d(y, x) . (Symmetry)
(3) d(x, y) ≤ d(x, z) + d(z, y) . (Triangle inequality)
d is called a metric on X . d(x, y) is sometimes called the distance between

x and y .
A sequence xk ∈ X is said to converge to x ∈ X if lim d(xk , x) = 0 ,
k→∞
that is, given > 0 , there exists N > 0 so that d(xk , x) < for all k ≥ N .
A set U ⊆ X is said to be open in X if for all x ∈ U , there is an r > 0
so that the set
B(x, r) ≡ {y ∈ X | d(x, y) < r}
is a subset of U (that is, B(x, r) ⊆ U ). A set C ⊆ X is said to be closed in

X if X − C is open in X . A set K ⊆ X is said to be compact if every open
covering {Uα }α∈I of K possesses a finite subcovering {Uαk }K
k=1 .
The collection of all open sets in X is called the topology induced by
the metric d . In general, one can specify a collection T of subsets of X as
a topology on S if the following conditions hold:
411
412 APPENDIX A. A BRIEF REVIEW ON ANALYSIS
(1) ∅ ∈ T , X ∈ T ;
S
(2) If Uα ∈ T for α ∈ I , then Uα ∈ T ;
α∈I
K
T
(3) If Uk ∈ T for k = 1, 2, · · · , K , then Uk ∈ T .
k=1
The sets belonging to T are called open sets of X .
A.2 Linear spaces (Vector spaces)
A linear space consists of the following:
(1) a field of scalars F (usually R or C);
(2) a set X of objects, called vectors;
(3) a rule, called vector addition, which associates with each pair of vectors
x , y ∈ X , called the sum of x and y , in such a way that
(a) addition is commutative, that is, x + y = y + x ;
(b) addition is associative, that is, x + (y + z) = (x + y) + z ;
(c) there is a unique vector 0 in X , called the zero vector, such that
x + 0 = x for all x ∈ X ;
(d) for each vector x in X there is a unique y (which is usually denoted

by −x) in X such that x + y = 0 ;
(4) a rule, called scalar multiplication, which associates with each scalar c
in F and vector x in X a vector cx in X , called the product of c and
x , in such a way that
(a) 1x = x for every x ∈ X ;
(b) (c1 c2 )x = c1 (c2 x) ;
(c) c(x + y) = cx + cy ;
(d) (c1 + c2 )x = c1 x + c2 x .
The linear space X with field F is often called “a linear space X over F”.
A.3. NORMED SPACES 413
A.3 Normed spaces
A vector space X is said to be a normed space if to every x ∈ X there is

associated a non-negative reall number kxk , called the norm of x , in such
a way that
(1) kx + yk ≤ kxk + kyk for all x and y in X ;
(2) kcxk = |c|kxk if x ∈ X and c is a scalar;
(3) kxk > 0 if x 6= 0 .
A metric d on X can be induced by the norm k · k: d(x, y) = kx − yk .
A.4 Completeness
(1) Cauchy sequence: Let (X, d) be a metric space. A sequence {xk }∞

k=1 ∈
X is called a Cauchy sequence if for all ε > 0 , there exists a number
N > 0 so that d(xi , xj ) < ε for all i , j ≥ N .
(2) Complete metric spaces: A metric space (X, d) is complete if all

Cauchy sequence has a limit in X , that is, if {xk }∞
k=1 is a Cauchy
sequence, there is a point x ∈ X so that d(xk , x) → 0 as k → ∞ .
A.5 Completion of a metric space
Let (X, d) be a metric space but not complete. There exists a complete
˜ such that
metric space (X̃, d)
(1) there is an one-to-one map ι : X → X̃ ;
˜
(2) d(ι(x), ι(y)) = d(x, y) ; (isometry)
(3) ι(X) is dense in X̃ .
We can view X̃ as all possible limit points of (X, d) , and d˜ is “the” extension
of d .
If x ∈ X , we also use the notation x ∈ X̃ to denote that ι(x) ∈ X̃ . That
is, we identify x as an element in X̃ . However, we have to have in mind
that conceptually this is not correct since every element in X̃ should be an

equivalent class of Cauchy sequence.
A complete normed linear space is called a Banach space.
A.6 Bounded linear maps between normed linear

spaces
Let X and Y be two linear spaces, and T : X → Y be linear, that is,
T (ax1 + bx2 ) = aT x1 + bT x2 ∀ a, b ∈ R(or C), x1 , x2 ∈ X .
If T : X → Y is ont-to-one and onto, then we can define the inverse map

T −1 : Y → X by T −1 y = x if and only if T x = y , so that T T −1 = T −1 T =
Id .
Definition A.1. Let (X, k · kX ) and (Y, k · kY ) be two normed linear spaces.
A linear map T : X → Y is bounded if there is a constant M ≥ 0 such that
kT xkY ≤ M kxkX ∀x ∈ X .
If no such constant exists, then we say that T is unbounded. If T : X → Y

is a bounded linear map, we define the operator norm of T by
kT k = inf{M | kT xkY ≤ M kxkX ∀ x ∈ X} .
We denote the set of all linear maps T : X → Y by L(X, Y ) , and the set
of all bounded linear maps T : X → Y by B(X, Y ) . When the domain and
range spaces are the same, we write L(X, X) = L(X) and B(X, X) = B(X) .
If in particular Y = R or C , then T is called a linear functional.
Note that
kT xkY
kT k = sup = sup kT xkY = sup kT xkY .
x6=0 kxkX kxkX ≤1 kxkX =1
Theorem A.2. A linear map is bounded if and only if it is continuous.
Definition A.3. Let X be a linear space. Two norms k · k1 and k · k2 on

X are equivalent if there are positive constants c and C such that
ckxk1 ≤ kxk2 ≤ Ckxk2 ∀x ∈ X .

A.7. CONVOLUTIONS AND INTEGRAL OPERATORS 415
Theorem A.4. Two norms on a linear space generate the same topology if
and only if they are equivalent.
Theorem A.5. All norms on Rn are equivalent.
Theorem A.6 (Open mapping theorem). Suppose that T : X → Y is

a one-to-one, onto bounded linear map between Banach space X and Y .
Then T −1 : Y → X is bounded.
A.7 Convolutions and integral operators
Suppose that f and g are two measurable functions on Rn . The convolution

of f and g is defined by
Z
(f ∗ g)(x) = f (y)g(x − y)dy x ∈ Rn .
Rn
Note that f ∗ g = g ∗ f , and the Tonelli theorem implies
kf ∗ gkL1 (Rn ) ≤ kf kL1 (Rn ) kgkL1 (Rn ) .
Theorem A.7 (Young’s inequality). Let p and q satisfy 1 ≤ p, q ≤ ∞ and

1 1 1 1 1
+ ≥ 1 , and let r be defined by = + − 1 . If f ∈ Lp (Rn ) and
p q r p q
g ∈ Lq (Rn ) , then f ∗ g ∈ Lr (Rn ) and
kf ∗ gkLr (Rn ) ≤ kf kLp (Rn ) kgkLq (Rn ) .
Proof. Use the identity

p p q q
|f (y)||g(x − y)| = |f (y)|1− r |f (y)| r |g(x−y)| r |g(x − y)|1− r
rp rq
and apply L r−p − Lr − L r−q type Hölder’s inequality.
If u : Rn → R satisfies certain integrability conditions, then we define

the operator K acting on the function u as follows:
Z
(Ku)(x) = k(x, y)u(y)dy ,
Rn
where k : Rn × Rn → R is called the integral kernel. Given a function

f ∈ Lp (Rn ) , define an integral kernel k(x, y) = f (x − y) , then by Young’s
inequality, the corresponding integral operator K ∈ B(Lq (Rn ), Lr (Rn )) .
A.8 The dual space and weak topology
A linear functional on a linear space X is a linear map from X to its scalar

fields. A continuous linear functional on a linear topological space (X ,T )
(with T denoting the topology which is called the strong topology) is a
continuous linear map from X to its scalar fields. Note that when talking
about the continuous linear functionals, we must have a strong topology to
begin with. The collection of continuous linear functionals on X is a linear
space, and is denoted by X 0 . X 0 is also called the dual space of X .
Example A.8 (Linear functional which is not continuous). Let X = C ∞ ([0, 1])
with sup-norm k · k∞ . Define the linear functional T : X → R by
T (f ) = f 0 (0) .
sin(n2 x)
Then T is not continuous. In fact, let fn (x) = n . Then fn → 0 in
(X, k · k∞ ) as n → ∞ , but T (fn ) 6→ 0 as n → ∞ .
The weak topology of X is the weakest/smallest topology on X so that

every element in X 0 is still continuous. The term “weak” is in the sense of
comparing to the “strong” topology. The weak topology of X is denoted by
Tw .
We can define a norm on X 0 as kf kX 0 = sup |f (x)| . Since X 0 with
kxkX ≤1
the topology induced by the dual space norm is a topological vector space,
we can talk about its dual space and weak topology. Every element x in X
can be considered as a continuous linear map which is defined by
x(f ) = f (x) .
In general, the dual space of X 0 is larger than X . For a reflexive space X ,

X 00 = X . The weak* topology of X 0 is the weakest/smallest topology on
X 0 so that every element in X is still continuous.
Theorem A.9 (Characteristics of weak and weak* convergence). Let (X, T )

be a topological vector space.
(1) A sequence {xk }∞

k=1 converges to x in the weak topology of X (this is
the same as saying that xk converges to x weakly in X), denoted by
xk * x in X , if and only if lim f (xk ) = f (x) for all f ∈ X 0 .
k→∞
A.9. BOUNDED LINEAR FUNCTIONAL ON LP (Ω) 417
(2) A sequence {fk }∞ 0

k=1 converges to f in the weak* topology of X , de-
∗
noted by fk * f in X 0 , if and only if lim fk (x) = f (x) for all x ∈ X .
k→∞
Using the notation from the linear algebra, from now on the value of the
linear map A at x is denoted by Ax instead of A(x) .
Theorem A.10 (Banach-Alaoglu). If V is an open set containing 0 in a

topological vector space X and if
K = {Λ ∈ X 0 | |Λx| ≤ 1 ∀ x ∈ V } ,
then K is weak*-compact in X 0 .
A.9 Bounded linear functional on Lp (Ω)

p
Theorem A.11. Let p ∈ (1, ∞] , q = p−1 . For g ∈ Lq (Ω) , define Fg :
Lp (Ω) → R as
Z
Fg (f ) = f (x)g(x)dx .
Ω
Then Fg is a continuous linear functional on Lp (Ω) with operator norm

kFg kLp (Ω)0 = kgkLq (Ω) .
Theorem A.12 (Riesz Representation). Suppose that 1 ≤ p < ∞ and

p
φ ∈ Lp (Ω)0 . Then there exists g ∈ Lq (Ω) , q = p−1 such that
Z
φ(f ) = f (x)g(x)dx ∀ f ∈ Lp (Ω)
Ω
and kφkLp (Ω)0 = kgkLq (Ω) .
Corollary A.13. For p ∈ (1, ∞) , the space Lp (Ω) is reflexive, that is,
Lp (Ω)00 = Lp (Ω) .
Corollary A.14. If {fk }∞ p

k=1 is a bounded sequence in L (Ω) for some p ∈
(1, ∞) , then there exists a subsequence {fkj }∞ p
j=1 such that fkj * f in L (Ω) .
Corollary A.15. If {fk }∞ ∞

k=1 is a bounded sequence in L (Ω) , then there
∗
exists a subsequence {fkj }∞ ∞
j=1 such that fkj * f in L (Ω) .
Theorem A.16. If fk * f in Lp (Ω) as k → ∞ , then
kf kLp (Ω) ≤ lim inf kfk kLp (Ω) .

k→∞
Theorem A.17. If fk * f in Lp (Ω) as k → ∞ , then {fk }∞

k=1 is bounded
in Lp (Ω) .
Theorem A.18. Suppose that Ω ⊆ Rn is bounded, and
sup kfk kLp (Ω) ≤ M < ∞ and fk → f a.e.

k
If 1 < p < ∞ , then fk * f in Lp (Ω) .

Appendix B
Important Topics in
Functional Analysis
B.1 The Hahn-Banach theorem
Definition B.1. A vector space X is said to be a topological vector space if

there is a topology τ on X so such that
(a) every point of X is a closed set, and
(b) the vector space operations (addition of vectors and multiplication with
scalars) are continuous with respect to τ .
Definition B.2. The dual space of a topological vector space X is the vector
space X 0 whose elements are the continuous linear functionals on X .
Proposition B.3. A complex-linear functional on X is in X 0 if and only if

its real part is continuous, and that every continuous real-linear u : X → R
is the real part of a unique f ∈ X 0 .
Definition B.4. A map p from a real vector space V to R ∪ {±∞} is said

to be sub-linear over V if
p(λu) = λp(u) ∀ u ∈ V ,λ > 0,

p(u + v) ≤ p(u) + p(v) ∀ (u, v) ∈ V × V .
Theorem B.5. Let X be a real vector space, p a sub-linear function over

X , M a vector subspace of X . Suppose that T a linear functional over M
419
420APPENDIX B. IMPORTANT TOPICS IN FUNCTIONAL ANALYSIS
and T x ≤ p(x) on M . Then there exists a linear functional T̃ over X such

that
T̃ x = T x ∀x∈M,
and
−p(−x) ≤ T̃ x ≤ p(x) ∀ x∈X.
Corollary B.6. If X is a normed space and x0 ∈ X , there exists T ∈ X 0

such that
T x0 = kx0 kX and |T x| ≤ kxkX ∀ x∈X.
Theorem B.7. Let A and B are disjoint, non-empty, convex sets in a

topological vector space X .
(a) If A is open, then there exists T ∈ X 0 and γ ∈ R such that
Re T x < γ ≤ Re T y
for every x ∈ A and every y ∈ B .
(b) If A is compact, B is closed, and X is locally convex, then there exist

T ∈ X 0 , γ1 , γ2 ∈ R such that
Re T x < γ1 < γ2 < Re T y
for every x ∈ A and every y ∈ B .
Theorem B.8. Suppose M is a subspace of a locally convex space X , and

x0 ∈ X . If x0 is not in the closure of M , then there exists T ∈ X 0 such
that T x0 = 1 but T x = 0 for every x ∈ M .
Theorem B.9. If f is continuous linear functional on a subspace M of a

locally convex space X , then there exists T ∈ X 0 such that T = f on M .
Theorem B.10. Suppose B is a convex, balanced, closed set in a locally

convex space X , x0 ∈ X , but x0 6∈ B . Then there exists T ∈ X 0 such that
|T x| ≤ 1 for all x ∈ B , but T x0 > 1 .
B.2. THE OPEN MAPPING AND CLOSED GRAPH THEOREM 421
B.2 The open mapping and closed graph theorem
Theorem B.11 (The Baire Category Theorem). Let X be a complete metric

space.
∞
(a) If {Un }∞
T
n=1 is a sequence of open dense subsets of X , then Un is
n=1
dense in X .
(b) X is not a countable union of nowhere dense sets.
Definition B.12 (Open mapping). Let X and Y be two topological vector

spaces. A mapping f : X → Y is said to be open if f (U ) is open in Y
whenever U is open in X .
Theorem B.13 (The Open Mapping Theorem). Suppose that X and Y be

Banach spaces, and T ∈ B(X, Y ) is surjective (i.e., onto). Then T is an
open mapping.
Corollary B.14 (The Bounded Inverse Theorem). Suppose that X and Y

be Banach spaces, and T ∈ B(X, Y ) is bijective (i.e., one-to-one and onto),
then the inverse map of T is bounded, or T −1 ∈ B(Y, X) . Equivalently,
there exist positive real numbers c and C such that
ckxkX ≤ kT xkY ≤ CkxkX ∀ x∈X.
Theorem B.15 (The Closed Graph Theorem). Suppose that X and Y are
Banach spaces, and T : X → Y is linear. If G = {(x, T x) | x ∈ X} is closed
in X × Y , then T ∈ B(X, Y ) .
B.3 Compact operators
Definition B.16 (Compact operators). Suppose X and Y are Banach

spaces and U is the open unit ball in X . A linear map T : X → Y is
said to be compact if the closure of T (U ) is compact in Y . It is clear that
T is then bounded. Thus T ∈ B(X, Y ) .
Definition B.17. An operator T ∈ B(X) is said to be invertible if there

exists S ∈ B(X) such that
ST = I = T S .
In this case, we write S = T −1 .
Definition B.18 (Spectrum and resolvent set). The specturm σ(T ) of an

operator T ∈ B(X) is the set of all scalars λ such that T −λI is not invertible,
and the resolvent set ρ(T ) is the complement of σ(T ) in the scalar field. Thus
λ ∈ σ(T ) if and only if at least one of the following two statements is true:
(i) The range of T − λI is not all of X .
(ii) T − λI is not one-to-one.
Definition B.19 (Classification of σ(T )). The spectrum of T ∈ B(X) is the

(disjoint) union of the following three sets:
(i) The point spectrum σp (T ) = {λ ∈ C | T − λI is not one-to-one} . If

λ ∈ σp (T ) , λ is also called an eigenvalue of T .
(ii) The continuous spectrum
σc (T ) = {λ ∈ C | T − λI is one-to-one, and has dense range}.
(iii) The residual spectrum
σr (T ) = {λ ∈ C | T − λI is one-to-one, and does not have dense range}.
Proposition B.20. The spectrum of a bounded operator T ∈ B(X) is

bounded.
Theorem B.21. Let X and Y be Banach spaces.
(a) If T ∈ B(X, Y ) and dim R(T ) < ∞ , then T is compact.
(b) If T ∈ B(X, Y ) , T is compact, and R(T ) is closed, then dim R(T ) <
∞.
(c) The compact operators form a closed subspace of B(X, Y ) in its norm-
topology.
(d) If T ∈ B(X) , T is compact, and λ 6= 0 , then dim N (T − λI) < ∞ .
(e) If dim X = ∞ , T ∈ B(X) , and T is compact, then 0 ∈ σ(T ) .

B.3. COMPACT OPERATORS 423
(f) If S ∈ B(X) , T ∈ B(X) , and T is compact, so are ST and T S .
Proof. (a) and (f) are trivial and left as exercises.
(b) If Y ≡ R(T ) is closed, then Y is complete, so that T is an open

mapping of X onto R(X) . Let U be the unit ball in X , then V ≡ T U
is open in Y . Since T is compact, V is pre-compact. Therefore, there
exist y1 , · · · , ym such that
m
[ 1
V ⊆ (yj + V ) . (1)
2
j=1
Let Z be the vector space spanned by y1 , · · · , ym . Then dim Z ≤ m ,

and Z is a closed subspace of Y . We also note that (1) implies V ⊆
Z + 21 V . Since Z = λZ for all λ 6= 0 ,
1 1 1
V ⊆Z + V ⊆Z +Z + V =Z + V .
2 4 4
We then see that

∞
\
V ⊆ (Z + 2−n V ) = Z .
n=1
However, it would further implies that kV ⊆ Z for all k ∈ N , so

Z=Y .
(c) Let Σ be the collection of compact operators in B(X, Y ) , U be the

unit ball in X , and T ∈ Σ . For every r > 0 , there exists S ∈ Σ with
kS − T kB(X,Y ) < r . Since SU is totally bounded, there exists points
x1 , · · · , xn in U such that SU is covered by the balls B(Sxi , r) . Since
kSx − T xkY ≤ r for every x ∈ U , it follows that T U is covered by the
balls of B(T xi , 3r) . Thus T U is totally bounded as well, so T ∈ Σ .
(d) Let Y = N (T − λI) . The restriction of T to Y is a compact operator

whose range is Y . By (b), dim(Y ) < ∞ .
(e) T : X → X cannot be an onto map since if it is onto, then dim R(T ) =

∞ which contradicts to (b) .
Definition B.22 (Adjoint operators). The adjoint operator T ∗ of an op-

erator T ∈ B(X, Y ) is the unique bounded operator belonging to B(Y 0 , X 0 )
satisfying
hT x, y ∗ iY = hx, T ∗ y ∗ iX .
Theorem B.23. Suppose X and Y are Banach spaces and T ∈ B(X, Y ) .

Then T is compact if and only if T ∗ is compact.
Proof. (⇒) Suppose T is compact. Let {yn∗ }∞

n=1 be a sequence in the unit
ball of Y 0 . Define
fn (y) = hy, yn∗ iY ∀y∈Y .
Since |fn (y1 )−fn (y2 )| ≤ ky1 −y2 kY , {fn }∞

n=1 is equi-continuous. Since T (U )
has compact closure in Y (as before, U is the unit ball of X), Arzela-Ascoli
theorem implies that {fn }∞ ∞
n=1 has a subsequence {fnj }j=1 that converges
uniformly on T (U ) . Since
kT ∗ yn∗ i − T ∗ yn∗ j kX 0 = sup |hT x, yn∗ i − yn∗ j iY | = sup |fni (T x) − fnj (T x)| ,
x∈U x∈U
the completeness of X 0 implies that {T ∗ yn∗ j }∞ ∗

j=1 converges. Hence T is
compact.
(⇐) can be proved in the same fashion.
Definition B.24. Suppose M is a closed subspace of a topological vector

space X . If there exists a closed subspace N of X such that
X =M +N and M ∩ N = {0} ,
then M is said to be complemented in X . In this case, X is said to be the

direct sum of M and N , and the notation X = M ⊕ N is used.
Lemma B.25. Let M be a closed subspace of a Banach space X .
(a) If dim M < ∞ , then M is complemented in X .
(b) If dim(X/M ) < ∞ , then M is complemented in X .
The dimension of X/M is also called the codimension of M in X .

Proof. Note that the closedness of M is only used in (b), while in (a)
the closedness is implied by the finite dimensionality (so no assumption
is needed).
(a) Let {e1 , · · · , en } be a basis for M . Then every x ∈ M has a unique

representation
x = α1 (x)e1 + · · · + αn (x)en .
αi is a continuous linear functional which vanishes on the span of

{e1 , · · · , ei−1 , ei+1 , · · · , en } , and can be extended to a continuous lin-
ear functional that only take non-zero values in the 1-dimensional
space spanned by ei . Let N be the intersection of the null space
of these extensions. Then X = M ⊕ N .
(b) Let {e1 , · · · , en } be a basis of X/M (closedness of M is used to define

the quotient space), and π : X → X/M be the quotient map. Pick
xi ∈ X so that πxi = ei , and define N to be the span of {x1 , · · · , xn } .
Then X = M ⊕ N .
Lemma B.26. Let M be a subspace of a normed space X . If M is not

dense in X , and if r > 1 , then there exists x ∈ X such that
kxkX < r but kx − ykX ≥ 1 ∀ y ∈ M.
Proof. There exists x1 ∈ X whose distance from M is 1, that is,
inf{kx1 − ykX |y ∈ M } = 1 .
Choose y1 ∈ M such that kx1 − y1 kX < r , and put x = x1 − y1 .
Theorem B.27. If X is a Banach space, T ∈ B(X) , T is compact, and

λ 6= 0 , then T − λI has closed range.
Proof. By (d) of Theorem B.21, dim N (T − λI) < ∞ . By (a) of Lemma

B.25, X is the direct sum of N (T − λI) and a closed subspace M . Define
an operator S ∈ B(M, X) by
Sx = T x − λx .
Then S is one-to-one on M . Also, R(S) = R(T − λI) . Similar to the proof

of Lax-Milgram theorem 5.15, to show that R(S) is closed, it suffices to show
the existence of an r > 0 such that
rkxkX ≤ kSxkX ∀x∈M.
Suppose the contrary that for every r > 0 , there exists {xn } in M such that
kxn kX = 1 , Sxn → 0 , and (after passage to a subsequence) T xn → x0 for
some x0 ∈ X (by the compactness of T ). It follows that λxn → x0 . Thus
x0 ∈ M since M is a closed subspace, and
Sx0 = lim (λSxn ) = 0 .

n→∞
Since S is one-to-one, x0 = 0 . However, kxn kX = 1 for all n , and x0 =

lim λxn , hence kx0 kX = |λ| > 0 .
n→∞
Corollary B.28. The continuous spectrum of a compact operator T ∈ B(X)

contains at most one point, namely 0.
Theorem B.29. Suppose X is a Banach space, T ∈ B(X) , T is compact,

r > 0 , and E is a set of eigenvalues λ of T such that |λ| > r . Then
(a) for each λ ∈ E , R(T − λI) 6= X , and
(b) E is a finite set.
Proof. We first show that either (a) or (b) is false then there exist closed
subspaces Mn of X and scalars λn ∈ E such that
M1 ( M2 ( M3 ( · · · , (1)
T (Mn ) ⊆ Mn for n ≥ 1 , (2)
(T − λn I)(Mn ) ⊆ Mn−1 for n ≥ 2 . (3)
Suppose (a) is false. Then R(T − λ0 I) = X for some λ0 ∈ E . Let

S = T − λ0 I , and define Mn = N (S n ) , i.e., the null space of S n . Since λ0
is an eigenvalue of T , there exists x1 ∈ M1 , x1 6= 0 . Since R(S) = X , there
is a sequence {xn }∞
n=1 in X such that Sxn+1 = xn , n = 1, 2, 3, · · · . Then
S n xn+1 = x1 6= 0 but S n+1 xn+1 = Sx1 = 0 .

Hence Mn is a proper closed subspace of Mn+1 . It follows that (1) to (3)

hold, with λn = λ0 .
Suppose (b) is false. Then E contains a sequence {λn } of distinct eigen-
values of T . Choose corresponding eigenvectors en , and let Mn be the
(finite-dimensional, hence closed) subspace of X spanned by {e1 , · · · , en } .
Since λn are distinct, {e1 , · · · , en } is a linearly independent set, so that
Mn−1 is a proper subspace of Mn . This gives (1). If x ∈ Mn , then
x = α1 e1 + · · · αn en ,
which shows that T x ∈ Mn and
(T − λn I)x = α1 (λ1 − λn )e1 + · · · + αn−1 (λn−1 − λn )en−1 ∈ Mn−1 .
Thus (2) and (3) hold.

Once we have closed subspace Mn satisfying (1) to (3), Lemma B.26
gives us vectors yn ∈ Mn , for n = 2, 3, 4, · · · , such that
kyn kX ≤ 2 and kyn − xkX ≥ 1 if x ∈ Mn−1 . (4)
If 2 ≤ m < n , define
z = T ym − (T − λn I)yn .
By (2) and (3), z ∈ Mn−1 . Hence (4) shows that
kT yn − T ym kX = kλn yn − zkX = |λn |kyn − λ−1

n zkX ≥ |λn | > r .
The sequence {T yn }∞
n=1 has therefore no convergent subsequences, although
{yn }∞
n=1 is bounded, contradicting to the compactness of T .
Remark B.30. Let H denote a Hilbert space. T ∈ B(H) is said to be

normal if T T ∗ = T ∗ T . A much deeper result states that a normal operator
T ∈ B(H) is compact if and only if it satisfies the following two conditions:
(a) σ(T ) has no limit point except possibly 0 .
(b) If λ 6= 0 , then dim N (T − λI) < ∞ .

Theorem B.31 (The Fredholm Alternative). Suppose X is a Banach space,

T ∈ B(X) , and T is compact.
(a) If λ 6= 0 , then the four numbers
α = dim N (T − λI) α∗ = dim N (T ∗ − λI)

β = dim X/R(T − λI) β ∗ = dim X/R(T ∗ − λI)
are equal and finite.
(b) If λ 6= 0 and λ ∈ σ(T ) , then λ is an eigenvalue of T and of T ∗ .
(c) σ(T ) is compact, at most countable, and has at most one limit point,
namely, 0.
Proof. Suppose M0 is a closed subspace of a locally convex space Y , and k

is a positive integer such that k ≤ dim Y /M0 . Then there are vectors y1 ,
· · · , yk in Y such that the vector space Mi generated by M0 and y1 , · · · , yi
contains Mi−1 as a proper subspace. Each Mi is closed, and hence by The-
orem B.8, there are continuous linear functionals T1 , · · · , Tk on Y such
that Ti yi = 1 but Ti y = 0 for all y ∈ Mi−1 . These functionals are linearly
independent, so if Σ denotes the space of all continuous linear functionals
on Y that annihilate M0 , then
dim Y /M0 ≤ dim Σ .
Let S = T − λI . Apply this with Y = X , M0 = R(S) . Since R(S)

is closed, Σ = R(S)⊥ = N (S ∗ ) , so β ≤ α∗ . Next, take Y = X 0 with its
weak*-topology, and M0 = R(S ∗ ) . A result from functional analysis states
that R(S ∗ ) is weak*-closed. Since Σ consists of all weak*-continuous linear
functional on X 0 that annihilate R(S ∗ ) , Σ is isomorphic to ⊥ R(S ∗ ) = N (S) ,
hence β ∗ ≤ α .
Next we show that α ≤ β , and the same proof can be used to show that
α∗ ≤ β ∗ , so the proof of (a) (and hence (b) and (c)) is complete. Assume
the contrary that α > β . By (d) of Theorem B.21, α < ∞ . By Lemma
B.25, there exists closed subspaces E and F such that dim F = β and
X = N (S) ⊕ E = R(S) ⊕ F .
Every x ∈ X has unique representation x = x1 + x2 , with x1 ∈ N (S) ,

x2 ∈ E . Define π : X → N (S) by πx = x1 . It is easy to see (by the closed
graph theorem B.15) that π is continuous.
Since we assume that dim N (S) > dim F , there is a linear mapping φ of
N (S) onto F such that φx0 = 0 for some x0 6= 0 . Define
Φx = T x + φπx ∀ x∈X.
Then Φ ∈ B(X) . Since dim R(φπ) < ∞ , φπ is a compact operator, hence

so is Φ .
Observe that Φ − λI = S + φπ . If x ∈ E , then πx = 0 , so (Φ − λI)x =
Sx ; hence
(Φ − λI)(E) = R(S) .
If x ∈ N (S) , then πx = x , (Φ − λI)x = φx ; hence
(Φ − λI)(N (S)) = φ(N (S)) = F .
Therefore, R(Φ − λI) = R(S) + F = X . Moreover, λ is an eigenvalue of Φ

(with x0 as a corresponding eigenvector), and since Φ is compact, Theorem
B.29 states that R(Φ − λI) cannot be all of X .
Corollary B.32. The residual spectrum of a compact operator T ∈ B(X)

contains at most one point, namely 0 . Moreover, σ(T ) = σp (T ) ∪ {0} .
Corollary B.33. Suppose that H is a Hilbert space, and T ∈ B(H) is

compact. Then R(T − λI) = N (T ∗ − λI)⊥ for all λ 6= 0 .
Remark B.34. If T ∈ B(X) is compact, then the injectivity of T − λI

implies the invertibility of T − λI if λ 6= 0 .
Remark B.35. A much deeper result states that the spectrum of a bounded
operator T ∈ B(X) is also compact.
B.3.1 Symmetric operators on Hilbert Spaces
Let H be a Hilbert space, and T ∈ B(H) . By Riesz representation theorem,

given a continuous linear functional y ∗ ∈ H0 , there exists y ∈ H such that
hh, y ∗ iH = (h, y)H ∀ h ∈ H.
In particular, let h = T x , and suppose the representation of T ∗ y ∗ is z , then
(x, z)H = hx, T ∗ y ∗ iH = hT x, y ∗ iH = (T x, y)H ∀ h ∈ H.
The element z ∈ H is denoted by T 0 y . In this case, T 0 is also called the

adjoint operator of T (and T 0 can be thought as the representation of T ∗ ).
Definition B.36 (Symmetry). The operator T ∈ B(H) is called symmetric

if T = T 0 .
Lemma B.37. Suppose that T ∈ B(H) be symmetric, and
m≡ inf (T u, u)H , M ≡ sup (T u, u)H .

kukH =1 kukH =1
Then σ(T ) ⊆ [m, M ] , and m, M ∈ σ(T ) .
Proof. Let λ > M . Then L : H → H0 defined by
hLu, ϕiH = (λu − T u, ϕ)H ∀ϕ∈H
is bounded and coercive: the boundedness is trivial, and the coercivity fol-
lows from that
hLu, uiH = (λu − T u, u)H ≥ (λ − M )kuk2H .
Therefore, by the Lax-Milgram theorem, L : H → H0 is one-to-one and onto,

so is λI − T since λI − T is the representation of L . Therefore, λ 6∈ σ(T ) .
Similarly, λ 6∈ σ(T ) if λ < m . So σ(T ) ⊆ [m, M ] .
Let [u, v] = (M u − T u, v)H . The proof of the Schwarz inequality (The-
orem 3.1) implies that
|[u, v]| ≤ |[u, u]|1/2 |[v, v]|1/2 .

Taking the supremum over all v such that kvkH = 1 , then

1/2
kM u − T ukH ≤ C(M u − T u, u)H ∀u∈H (1)
for some constant C .

Let {uk }∞
k=1 be such that kuk kH = 1 , and (T uk , uk )H → M . Then (1)
implies kM uk − T uk kH → 0 as k → ∞ . If M 6∈ σ(T ) , M I − T is invertible
and has a bounded inverse (by the bounded inverse theorem), so
uk = (M I − T )−1 (M uk − T uk ) → 0 in H
which contradicts to kuk kH = 1 for all k . Hence M ∈ σ(T ) . Similarly,

m ∈ σ(T ) .
Theorem B.38. Let H be a separable Hilbert space, and suppose that T ∈

B(H) is compact and symmetric. Then there exists a countable orthonormal
basis of H consisting of eigenvectors of T .
Proof. Let {λk }∞

k=1 be the sequence of distinct eigenvalues of T , λk 6= 0 .
Set λ0 = 0 , and Hk = N (T − λk I) for k ≥ 0 . Then dim Hk < ∞ if k > 0 .
Moreover, if xi ∈ Hi and xj ∈ Hj , then
λi (xi , xj )H = (T xi , xj )H = (xi , T xj )H = λj (xi , xj )H ⇒ (xi , xj )H = 0 if i 6= j .
Therefore, the subspaces Hi and Hj are orthogonal.

Let H̃ be the smallest subspace of H consisting of all these Hi , i =
0, 1, · · · . Then
m
nX o
H̃ = ck uk m ∈ N ∪ {0} , uk ∈ Hk , ak ∈ R .

k=0
We note that T (H̃) ⊆ H̃ , and this further implies that T (H̃⊥ ) ⊆ H̃⊥ since
(T u, v)H = (u, T v)H = 0 ∀ u ∈ H̃⊥ , v ∈ H̃ .
The operator T̃ ≡ T |H̃⊥ , the restriction of T to H̃⊥ , is also compact and

symmetric. In addition, σ(T̃ ) = {0} , since any nonzero eigenvalue of T̃
would be an eigenvalue of T as well. According to the previous lemma,
(T̃ u, u)H = 0 for all u ∈ H̃ ⊥ . But then if u, v ∈ H̃ ⊥ ,
2(T̃ u, v)H = (T̃ (u + v), (u + v))H − (T̃ u, u)H − (T̃ v, v)H = 0

Hence T̃ = 0 on H̃⊥ . As a consequence, H̃⊥ ⊆ N (T ) ⊆ H̃ , so H̃⊥ = {0} .

Thus H̃ is dense in H .
An orthonormal basis of H then can be obtained by choosing an or-
thonormal basis for each subspace Hk , k = 0, 1, · · · . Note that the sepa-
rability of H implies that H0 has a countable orthonormal basis, and these
basis vectors are all eigenvectors corresponding to λ0 = 0 .

218alecturenotes - 6 07 12

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

218alecturenotes - 6 07 12

Uploaded by

Copyright:

Available Formats

MAT218: Lecture Notes on Partial Differential

2 Introduction to Sobolev Spaces 11

2.1.12 Weak convergence . . . . . . . . . . . . . . . . . . . . 26

2.4.2 Fractional-order Sobolev spaces via difference quotient

3 Fourier Series and its Applications 119

4 The Laplace and Poisson Equations 155

4.1.1 Uniform and Hölder Continuous Functions . . . . . . 157

5 Second Order Linear Elliptic Equations 181

5.5.2 The Dirichlet problem . . . . . . . . . . . . . . . . . . 199

6 Linear Parabolic Equations 241

6.2.5 Higher regularity and compatibility conditions . . . . 259

7 Linear Hyperbolic Equations 270

8 More Topics on Elliptic Equations 332

9 Fluid dynamics 358

A A Brief Review on Analysis 411

A.5 Completion of a metric space . . . . . . . . . . . . . . . . . . 413

B Important Topics in Functional Analysis 419

Theorem 1.1 (Divergence Theorem). Let Ω ⊆ Rn be a Lipschitz do-

Now suppose that f is a scalar C 1 -function, and N = (N1 , · · · , Nn ) . By

Suppose v ∈ C 2 (Ω) as well. Interchanging u and v in (1.1) and forming the

1.2 Examples of PDEs

Let u be the concentration of some physical quantity (u = u(x, t)) in a

where q is the strength of sources for the quantity. If u is smooth, by the

for all Lipschitz domain U . We then obtain the equation of continuity

%t + div(ρu) = 0 ∀x ∈ Ω,t ∈ R. (1.3)

(a) F (u) = 0 if u = 0 or u > L ,

(b) F 0 (u) > 0 if u ∈ (0, umax ) , and

(c) F 0 (u) < 0 if u ∈ (umax , L) .

If F is differential, and F 0 (u) = c(u) . Then the equation of continuity

1.2.2 The heat equation

Let u(t, x) defined on (0, T ] × Ω be the temperature of a material body at

where N denotes the outward-pointing unit normal of U . Assume that u

Since U is arbitrary, the equation above implies

c(x)%(x)ut (t, x) − div(k(x)Du(t, x)) = 0 ∀ x ∈ Ω , t ∈ (0, T ].

If furthermore c and % are constants, then after rescaling of time we have

This is the standard heat equation (the prototype equation of parabolic

1. Initial condition: u(0, x) = ϕ(x) , ϕ(x) is a given function.

2. Boundary condition: if Ω is bounded, some boundary condition of u

(a) Dirichlet condition: u(t, x) = ψ(t, x) for all x ∈ ∂Ω , t ≥ 0 , where

1.2.3 The Laplace equation

The stationary solution (meaning that the solution is time independent) of

1.2.4 The 1-d wave equation

1. From Hooke’s law: imagine an array of little weights of mass m inter-

If the array of weights consists of N weights spaced evenly over the

∂2u KL2 u(x + h, t) − 2u(x, t) + u(x − h, t)

utt (x, t) = c2 uxx (x, t) . (1.6)

(2) Equation of vibrating string: let u(x, t) measure the distance of a

Assuming only motion in the vertical direction, the horizontal compo-

The difference of the vertical component of T1 and T2 induces the

If µ is the density of the string, then m = µh ; hence

∂2u T (x + h)ux (x + h, t) − T (x)ux (x, t)

Taking the limit h → 0 , we obtain

µutt (x, t) = [T (x)ux (x, t)]x . (1.7)

If there is an external forcing f acting on the string, then (1.7) becomes

µutt (x, t) = [T (x)ux (x, t)]x + f . (1.8)

If T is constant, then (1.8) reduces to

1. Vibration string with fixed ends: u(0, t) = u(L, t) = 0 .

1.2.5 Equilibrium of a membrane

Let T denote the tension of a membrane, and z = u(x1 , x2 ) be the membrane

The work done by the inner elastic force is equal to

In other words, to deform a membrane from its unforced equilibrium state

If u ∈ C 2 (Ω) , we may integrate by parts to obtain